markdown
stringclasses 3
values | metadata
dict |
|---|---|
## Retrieval-Augmented Generation with Graphs (GraphRAG)
**Haoyu Han[1][∗], Yu Wang[2][∗], Harry Shomer[1], Kai Guo[1], Jiayuan Ding[5], Yongjia Lei[2],**
**Mahantesh Halappanavar[3], Ryan A. Rossi[4], Subhabrata Mukherjee[5], Xianfeng Tang[6], Qi He[6],**
**Zhigang Hua[7], Bo Long[7], Tong Zhao[8], Neil Shah[8], Amin Javari[9], Yinglong Xia[7], Jiliang Tang[1]**
1Michigan State University, 2University of Oregon, 3Pacific Northwest National Laboratory
4Adobe Research, 5Hippocratic AI, 6Amazon, 7Meta,8Snap Inc.,,9The Home Depot,
```
{hanhaoy1, shomerha, guokai1, tangjili}@msu.edu,
{yuwang, yongjia}@uoregon.edu, hala@pnnl.gov, ryarossi@gmail.com,
{jiayuan, subho}@hippocraticai.com, {xianft, qih}@amazon.com,
{zhua, bolong, yxia}@meta.com, {tong, nshah}@snap.com, amin_javari@homedepot.com
### Abstract
```
Retrieval-augmented generation (RAG) is a powerful technique that enhances downstream task execution by retrieving additional information, such as knowledge,
skills, and tools from external sources. Graph, by its intrinsic "nodes connected by
edges" nature, encodes massive heterogeneous and relational information, making
it a golden resource for RAG in tremendous real-world applications. As a result,
we have recently witnessed increasing attention on equipping RAG with Graph, i.e.,
GraphRAG. However, unlike conventional RAG, where the retriever, generator, and
external data sources can be uniformly designed in the neural-embedding space, the
uniqueness of graph-structured data, such as diverse-formatted and domain-specific
relational knowledge, poses unique and significant challenges when designing
GraphRAG for different domains. Given the broad applicability, the associated
design challenges, and the recent surge in GraphRAG, a systematic and up-to-date
survey of its key concepts and techniques is urgently desired. Following this motivation, we present a comprehensive and up-to-date survey on GraphRAG. Our
survey first proposes a holistic GraphRAG framework by defining its key components, including query processor, retriever, organizer, generator, and data source.
Furthermore, recognizing that graphs in different domains exhibit distinct relational
patterns and require dedicated designs, we review GraphRAG techniques uniquely
tailored to each domain. Finally, we discuss research challenges and brainstorm
directions to inspire cross-disciplinary opportunities. Our survey repository is
[publicly maintained at https://github.com/Graph-RAG/GraphRAG/.](https://github.com/Graph-RAG/GraphRAG/)
### 1 Introduction
Retrieval-Augmented Generation (RAG), as a powerful technique to improve downstream tasks by
retrieving additional information from external data sources, has been successfully applied to various
real-world applications [87, 120, 514, 551]. In RAG frameworks, retrievers search for additional
knowledge, skills, and tools based on user-defined queries or task instructions. The retrieved content
is then refined by an organizer and seamlessly integrated with the original query or instruction, which
is further fed into the generator to produce the final answer. For example, when conducting questionanswering (QA) tasks, the classic "Retriever-then-Reader" frameworks [191, 196, 468, 562] retrieve
external factual knowledge to improve the answer faithfulness, which significantly benefit social
goodness and mitigate risks in high-stake scenarios (e.g., medical, legal, financial, and education
_∗Equal contribution._
Preprint. Under review.
-----
consultation [467, 472, 515]). Moreover, recent advancements in large language models (LLMs)
have further underscored the power of RAG in enhancing the social responsibility of LLMs, such as
mitigating hallucinations [397], enhancing interpretability and transparency [203], enabling dynamic
adaptability [360, 419], reducing privacy risks [512, 513], ensuring reliability/robust responses [105,
460], and promoting fair treatment [362].
Building on the unprecedented success of RAG and further considering the ubiquity of graphs in realworld applications [545], recent research has explored the integration of RAG with graph-structured
data. Unlike textual or visual data, graph-structured data encodes heterogeneous and relational
information through its intrinsic "nodes connected by edges" nature. For example, individuals
connected by social relationships of social networks usually exhibit homophily behaviors [291],
sequential decision-making steps in plans follow casual dependency [454], and atoms belonging to the
same functional group within a molecule possess unique structural properties [103, 508]. Designing
the RAG that utilizes relational information requires adapting its core components, such as the retriever
and generator, to seamlessly integrate graph-structured data, resulting in GraphRAG. Different from
RAG, which predominantly uses semantic/lexical similarity search [104, 120], GraphRAG offers
unique advantages in capturing relational knowledge by leveraging graph-based machine learning
(e.g., Graph Neural Networks (GNNs)) and graph/network analysis techniques (e.g., Graph Traversal
Search and Community Detection [98, 428]). For example, considering the query “What drugs are
used to treat epithelioid sarcoma and also affect the EZH2 gene product?" [452], blindly executing
the existing BM25 or embedding-based search that relies solely on semantic/lexical similarity ignores
relational knowledge encoded in graph structure. In contrast, some GraphRAG methods traverse the
graph along the relational path “Disease (Epithelioid Sarcoma) → [indication] → Drug ← [target]
_←_ Gene/Protein (EZH2 gene product)" to retrieve neighbors of Epithelioid Disease following the
relation [indication], neighbors of Gene EZH2 following the relation [target], and find their intersected
drug [186, 271, 428]. Moreover, some domains involve entities with extremely complex geometry that
require dedicated model design to characterize. For example, 3D structures in molecular graphs [52,
445] and hierarchical tree structures commonly found in product taxonomies (e.g., on Amazon [529]),
in document sections (e.g., when using Adobe Acrobat [537]), and social networks (e.g., at Snap [277])
requires carefully designed graph encoders (or, more precisely, geometric encoders) with appropriate
expressiveness to capture structural nuances [277, 527]. Simply verbalizing node texts and feeding
them into LLMs cannot express complex geometric information and becomes infeasible given the
exponentially growing textual descriptions as neighborhood layers expand.
Despite the above advantages of GraphRAGs over RAGs, designing appropriate GraphRAGs faces
unprecedented challenges due to the following differences in graph-structured data:
- Difference 1 - Unified versus (vs.) Diverse-Formatted Information: Unlike conventional
RAG, where semantic information can be uniformly represented as a 2D grid of image patches
or a 1D sequence of textual corpora, graph-structured data often encompass diverse formats
and are stored in heterogeneous sources [4, 26, 434]. For example, document graphs embed
entities as sentence chunks [98, 428], knowledge graphs store graph information as triplets or
paths [38], and molecule graphs consist of higher-order structures (e.g., cellular complexes) [26],
as shown in Figure 1. Some graph data may even be multimodal (e.g., text-attributed graphs
include both structural and textual attributes, and scene graphs combine structures and vision).
Consequently, this diversity necessitates different RAG designs. For retrievers, conventional RAG
assumes the target information is indexed in an image or text corpus, which can be uniformly
represented as vector embeddings and enable one-size-fits-all embedding-based retrieval. However,
retrievers for GraphRAG must consider the concrete format and source of the desired information,
making the one-size-fits-all design impractical. When dealing with knowledge graph questionanswering, information of nodes, edges, or subgraphs is usually fetched by graph search before
embedding matching-based retrieval [419, 492]. This fetching operation is usually conducted
by identifying relevant nodes/edges/subgraphs via entity linking, relational matching, and graph
search algorithms (e.g., Breadth-First Search, Depth-First Search, Monte Carlo Tree Search, and A*
search) [395, 419, 570], which is unachievable if solely through deep learning-based embedding
similarity search. Furthermore, the design of the retriever should ensure sufficient geometric
expressiveness to capture structural nuances. For instance, when retrieving APIs from a plan graph
to accomplish specific goals [355, 356, 454], it is essential to equip the retriever with directional
awareness. This enables the execution of APIs with resource dependencies in the correct order,
preventing conflicts and avoiding invalid operations. Similarly, designing expressive retrievers
capable of differentiating high-order subgraph structures, such as 6-cycle benzene versus vs. 4
-----
star Methane, and 3-star T-junction vs. 4-square road, is essential in drug design for disease
treatment [139] and road construction for city planning [209]. Beyond the retriever, the generator
also requires specialized designs. When retrieved content includes complex graph structures
with textual attributes, simply verbalizing the text of the subgraph and concatenating it into a
prompt may obscure critical structural information. In these cases, encoding the graph with
graph encoders such as GNNs before integrating it into generation can help preserve structural
nuances [134, 244, 434, 443, 456].
- Difference 2 - Independent vs. Interdependent Information: In conventional RAG, information
is stored and utilized independently. For example, documents are split into chunks, such as
individual sentences, paragraphs, or a fixed number of tokens, based on the document context
and downstream task [21, 562]. Each chunk is then indexed and stored independently in a vector
database. This independence prevents the retrieval from capturing chunk relations, which hinders
performance on tasks requiring multi-hop reasoning and long-form planning. However, GraphRAG
stores chunks as interconnected nodes with edges denoting their relations, which can benefit
retrieval, organization, and generation. For retrieval, these edges could enable multi-hop traversal
to capture other chunks that share a logical connection with existing retrieved chunks. Furthermore,
the retrieved content can be organized not only by their semantic meaning (e.g., reranking [43,
172, 256]) but also their structural relations (e.g., graph pruning [377, 431]). During the generation
phase, squeezing interdependency (e.g., positional encoding [361, 549]) to the generator would
encode richer structural signals into the generated content.
- Difference 3 - Domain Invariance vs. Domain-specific Information: The relations in graphstructured data are domain-specific. Unlike images and texts, where different domains often share
transferable semantics [254, 286], such as textures and grains in images or vocabulary defined by
the tokenizer in texts, graph-structured data lacks explicit transferable units. This shared basis
in images and texts lays the foundation for designing encoders with geometric invariance and
enables the well-known data-scaling law. However, for graph-structured data, the underlying data
generation process governing the generated graphs varies significantly across different domains.
This variability makes the relational information highly domain-specific, and it is nearly impossible
to design a unified GraphRAG applicable to different domains. For example, when predicting
the topic of an academic paper, the widely accepted homophily assumption suggests retrieving
references from the paper to inform its topic prediction [563]. However, this homophily assumption
is not suitable when classifying the role of an airport in a flight network, where hubs are often
sparsely distributed across a country with no direct connections [68]. Moreover, even within the
same graph from the same domain, different tasks may necessitate distinct GraphRAG designs.
For example, when designing an automatic email completion system to optimize communication
efficiency in a company, both content relevance and tone coherence should be considered [429].
To ensure the content relevance of the generated emails, one might assume that close emails (i.e.,
emails from the same conversation thread) share similar content and thus should be retrieved for
reference. However, to maintain tone coherence, emails from staff with similar roles might be
retrieved, even if they do not share close social relations (e.g., between subordinates and superiors)
but instead hold similar structural roles within the company (e.g., as managers of different teams).
Despite the above differences that have driven extensive research in GraphRAG, the current research
landscape in this field remains fragmented, with significant variation in concepts, techniques, and
datasets across studies. Moreover, current GraphRAG research primarily focuses on knowledge and
document graphs as surveyed in Figure 2, often overlooking broader applications in other domains
like infrastructure graphs. This imbalance not only hampers the advancement of GraphRAG but
also risks creating a "bubble effect" that restricts the scope of future exploration. To address these
challenges, we present a comprehensive and up-to-date review of GraphRAG, aiming to unify the
GraphRAG framework from the global perspective while also specializing its unique design for each
domain from the local perspective. The key contributions of this survey are as follows:
- A Holistic Framework of GraphRAG: We propose a holistic framework of GraphRAG consisting
of five key components: query processor, retriever, organizer, generator, and graph data source.
Within each component, we review representative GraphRAG techniques.
- Specialization of GraphRAG in different domains: We categorize GraphRAG designs into 10
distinct domains based on their specific applications, including knowledge graph, document
graph, scientific graph, social graph, planning & reasoning graph, tabular graph,
-----
Figure 1: RAG works on text and images, which can be uniformly formatted as 1D sequences or 2D
grids with no relational information. In contrast, GraphRAG works on graph-structured data, which
encompasses diverse formats and includes domain-specific relational information.
infrastructure graph, biological graph, scene graph, and random graph . For each domain,
we review their unique applications and specific graph construction methods. We then summarize
the distinctive designs of each component within our proposed holistic GraphRAG framework and
collect rich benchmark datasets and tool resources.
- Challenges and Future Directions: We highlight the challenges of current GraphRAG research
and pinpoint future opportunities for further advancing GraphRAG into the new frontier.
In the following, we highlight the differences between our
survey and existing surveys. Despite the urgent need for a systematic overview of GraphRAG, most existing surveys focus
on general RAG within the context of i.i.d. data [11, 120, 227,
551, 561]. Before the advent of LLMs, earlier surveys focused
on textual RAGs [11, 227]. With the recent unprecedented success achieved by foundational models such as LLMs, various
surveys have explored foundational-model-powered RAG in
different modalities. Gao et al. [120] group existing RAG ap- Figure 2: Publications of
proaches into three categories (Naive, Advanced, and Modular GraphRAGs in different doRAGs), summarize three core techniques (Retrieval, Genera- mains based on surveyed papers
tion, and Augmentation), and review evaluation metrics. In
parallel, Zhao et al. [551] review representative RAG systems according to their corresponding
application and data modality. [561] focuses on reviewing trustworthy concerns and techniques
of RAG. However, none of them have a dedicated focus on graph-structured data. To the best of
our knowledge, only one very recent study [319] has specifically surveyed RAG in the context of
graph-structured data. However, this work mainly focuses on reviewing techniques introduced by
graphs under the conventional RAG architecture without specializing in reviewing diverse relations
and technical designs for graphs across different domains. In contrast to its holistic review philosophy,
we recognize the inherent heterogeneity of graph-structured data and specialize our GraphRAG
review across different domains. Specifically, we uncover the fundamental task applications (when
to retrieve), graph construction methods and relational rationales (what to retrieve), and GraphRAG
techniques (how to retrieve) for each domain. In this way, our survey provides a comprehensive
overview of GraphRAG for information retrieval, data mining, and machine learning communities
and domain-specific insights that facilitate interdisciplinary research and industrial opportunities.
Our survey is structured as follows: Section 2 introduces the holistic framework of GraphRAG and
introduces representative techniques for its five key components. From Section 3 to 9, we delve
into specific domains, reviewing unique task applications, summarizing existing graph construction
methods that guide GraphRAG design for that domain, highlighting domain-specific techniques
for each of the five components within our proposed holistic framework, and presenting existing
GraphRAG resources (e.g., benchmark datasets and tools) used across different domains. Finally, we
discuss research challenges and opportunities in Section 10 and conclude our survey in Section 11.
-----
Figure 3: A holistic framework of GraphRAG and representative techniques for its key components.
### 2 A Holistic Framework of GraphRAG
Based on existing literature on GraphRAG, we present a holistic framework of GraphRAG. Next, we
introduce the basic problem setting and notation used throughout the whole framework.
**2.1** **Problem Setting and Notations**
Following the general setting of RAG, given a graph-structured data source G, the user-defined
query Q is further sent to query processor Ω[Processor] to obtain the pre-processed query _Q[ˆ]. After that,_
the retriever Ω[Retriever] retrieves the content C from the graph data source G based on _Q[ˆ]. Next, the_
retrieved content C is refined by the organizer Ω[Organizer] to formulate the refined content _C[ˆ]. Finally,_
the refined content _C[ˆ] triggers the generator Ω[Generator]_ to generate the final answer A. The above five
components are summarized as follows:
- Query Processor Ω[Processor]: Preprocessing the given query _Q[ˆ] = Ω[Processor](Q)._
- Graph Data Source G: Information organized in graph-structured format.
- Retriever Ω[Retriever]: Retrieve the content C = Ω[Retriever]( Q, G[ˆ] ) from G based on the query _Q[ˆ]._
- Organizer Ω[Organizer]: Arrange and refine the retrieved content _C[ˆ] = Ω[Organizer]( Q, C[ˆ]_ ).
- Generator: Generate answers A = Ω[Generator]( Q,[ˆ] _C[ˆ]) to answer query Q._
Unlike sequential-based textual data and grid-structured image data, graph-structured data encapsulates relational information. To effectively harness this relational information, the above five core
components of GraphRAG desire dedicated designs to handle graph-structured input/output and
support graph-based operations. For example, in the retriever component, conventional RAG in
the Natural Language Processing (NLP) utilizes sparse/dense encoders for index search [196, 468].
In contrast, GraphRAG employs graph traversal methods (e.g., entity linking and BFS/DFS) and
graph-based encoders (e.g., Graph Neural Networks (GNNs)) to produce embeddings for retrieval.
This motivates us to summarize key innovations and representative designs of GraphRAG for each of
the above five components under the holistic GraphRAG framework in the following.
**2.2** **Task Applications and Example Query Q**
Similar to the general RAG framework where the text-formatted query Q specifies the question
context or the task instruction. Query Q in GraphRAG could also be in the format of text. For
example, in knowledge graph-based question-answering, the query could be "What is the Capital
of China?" [272, 395]. In addition, the query could also be in other formats, such as smile strings
for molecular graphs [132], or could even be the combination of multiple formats, such as the scene
graph along with the text instruction [147]. Table 1 summarizes the common task applications and
exemplary queries used in each domain, as well as their representative references.
-----
Table 1: Summary of Task Applications and Exemplary Queries for GraphRAG in each domain.
-----
Figure 4: Existing techniques of query processor Ω[Processor] in GraphRAG.
Table 2: Difference of query processor Ω[Processor] between RAG and GraphRAG.
**Technique** **RAG** **GraphRAG**
Entity Recognition Extracting mentions in knowledge bases Extracting mentioned nodes in graphs.
Relational Extraction Extracting textual relations Extracting graph edge relations
Query Structuration Structuring text query to SQL, SPARQL Structuring text query to GQL
Query Decomposition Decomposed queries are separate Decomposed queries are logically related
Query Expansion Expansion based on semantic knowledge Expansion based on relational knowledge
**2.3** **Query Processor Ω[Processor]**
Unlike RAG, where both queries and data sources are purely text-formatted, data sources used in
GraphRAG are graph-structured, which raises challenges in bridging text-formatted queries and
graph-structured data sources. For example, the information that connects the knowledge graph
and the query "Who is Justin Bieber’s brother?" is not a specific passage but instead the entity
"Justin Bieber" and the relation "brother of". Many techniques are proposed to correctly extract this
information from the query, including entity recognition, relational extraction, query structuration,
query decomposition, and query expansion. In the following, we first review each of these five query
processing techniques within the broader NLP domain, followed by a focused examination of their
unique adaptations for GraphRAG.
**2.3.1** **Name Entity Recognition**
Named Entity Recognition (NER) aims to identify mentions of entities from the text that belong to
predefined categories, such as persons, locations, or organizations, and it serves as a fundamental
component for numerous natural language applications [160, 199, 228, 293]. NER techniques can
be broadly categorized into four main approaches: (1) rule-based methods, which rely entirely on
handcrafted rules and require no annotated data; (2) unsupervised learning methods, which use
unsupervised algorithms without labeled training examples; (3) feature-based supervised learning
methods, which depend on supervised algorithms and careful feature engineering; and (4) deep
learning approaches, which automatically discover representations needed after (un)-supervised
training the deep learning models. Recent LLMs fall into the category of deep learning approaches
and have demonstrated unprecedented success for NER. More details about these techniques and
their resources can be found in Li et al. [228].
Specifically, in the GraphRAG context, entity recognition primarily uses deep learning techniques
(e.g., EntityLinker [395, 493] and LLM-based extraction [186]) to identify entities in queries grounded
by nodes in the given graph data sources. This step is vital for applications such as knowledge graphbased question answering [395, 492, 493]. For example, given the question, "What is the best way to
guess the color of the eye of the baby?", NER extracts entities such as "baby", "eye", and "color",
which correspond to nodes in the knowledge graph and are treated as the seed nodes to initialize
the retrieval process thereafter [443, 347]. For more recent GraphRAG research, NER has evolved
beyond identifying the entity names but instead their structures. For example, Jin et al. [186] leverages
LLMs to recognize node types in the graph, which further guides the retriever to identify nodes that
-----
match the recognized types for next-round exploration. For example, given the question "Who are the
authors of ‘Language Models are Unsupervised Multi-task Learners’?" the initially recognized entity
should not only be based on the semantic name "Language Models are Unsupervised Multi-task
Learners" but also be based on the type of that entity, which is the paper node in this case. Accurately
recognizing the names and structures of entities in GraphRAG reduces cascading errors and provides
a solid foundation for subsequent retrieval and generation steps.
**2.3.2** **Relational Extraction**
Similar to NER, relational extraction (RE) is a long-standing technique in NLP to identify relations
among entities and is widely applied to structured search, sentiment analysis, question answering,
summarization, and knowledge graph construction [47, 230, 303]. Recent advances in RE have been
largely driven by deep learning techniques, and they can be summarized into three perspectives: text
representation, context encoding, and triplet prediction, more details of which can be found in Pawar
et al. [317], Nasar et al. [303], Han et al. [141].
For GraphRAG, RE serves two key purposes: constructing graph-structured data sources (e.g.,
knowledge graphs) by extracting triplets and matching the relations mentioned in the query and the
graph data source to guide the graph search. For instance, given a query like "What is the capital of
China?", relational extraction identifies the relation "capital of" and searches for corresponding edges
via vector similarity in the knowledge graph, which guides the neighborhood selection and graph
traversal direction [119, 200, 272, 273].
**2.3.3** **Query Structuration**
Query structuration transforms queries into formats tailored to specific data sources and tasks. It
often converts natural language queries into structured formats like SQL or SPARQL [181, 238]
to interact with relational databases. Recent advancements leverage pre-trained and fine-tuned
LLMs to generate structured queries from natural language input to query databases. For graphstructured data, Graph Query Language (GQL) has emerged, such as Cypher, GraphQL, and SPARQL,
which enables complex interactions with property graph databases. Additionally, Jin et al. [186]
introduced a technique that decomposes complex queries into multiple structured operations, including
node retrieval, feature fetching, neighbor checks, and degree assessment, enhancing precision and
adaptability in querying.
**2.3.4** **Query Decomposition**
Query decomposition [447] aims to split the input query into multiple distinct subqueries, which are
used to first retrieve sub-results and aggregate these sub-results together for the final results. In most
existing RAG and GraphRAG, decomposed queries usually possess explicit logic connections that
can handle complex tasks that require multistep reasoning and planning [248, 316, 355, 372, 477].
For example, a query like "Please generate an image where a girl is reading a book, and her pose is
the same as the boy in ‘example.jpg’ then describe the new image with your voice" involves multiple
subtasks [477], each of which would be completed by a specific sub-query. In addition, Park et al.
[316] enhance the decomposition of the query by building a question graph where each sub-query is
represented as a triplet within the graph. These graph-structured sub-queries effectively guide the
retriever/generator through multi-step promptings.
**2.3.5** **Query Expansion**
Query Expansion enriches a query by adding meaningful terms with similar significance [12], which
primarily addresses three challenges: (1) user-submitted queries are ambiguous and relate to multiple
topics; (2) queries may be too brief to fully capture user intent; and (3) users are often uncertain about
what they are seeking. Generally, it can be categorized into manual query expansion, automatic query
expansion, and interactive query expansion. More recently, LLM-based query expansion has been a
prominent area due to the creativity of the generated content[54, 173, 221]
Unlike existing methods that mostly focus on textual similarities and overlook relations, QE in
GraphRAG augments LLM expansion with structured relations. For example Xia et al. [459] expands
the query by leveraging neighboring nodes of the mentioned entities in the query. Alternatively, Wang
et al. [406] convert the query into several sub-queries using pre-defined templates.
-----
Figure 5: Visualizing representative retrievers used in GraphRAG.
Table 3: Categorizing representative retrievers used in GraphRAG.
**Method/Strategy** **Input** **Output** **Description**
Entity Linking Entity Mention Node Match query entity and graph node
Relational Matching Relation Mention Edge Match query relation and graph edge
Graph Traversal Node/Edge Graph Expand seed nodes/edges into subgraphs
Graph Kernel (Sub)Graph (Sub)Graph Match query graph and candidate graph
Shallow Embedding Any Any Embedding similarity match query and candidate
Deep Embedding Any Any Embedding similarity match query and candidate
Domain Expertise Expertise Rule Any Match Domain Expertise with nodes/edges/graphs
**2.4** **Retriever Ω[Retriever]**
After obtaining the processed query _Q[ˆ], the retriever Ω[Retriever]_ identifies and retrieves relevant content
_C from external graph sources G to augment the downstream task execution:_
_C = Ω[Retriever]( Q, G[ˆ]_ ) (1)
Recently, retrievers have been increasingly integrated with LLMs to mitigate hallucination issues [397], address privacy concerns [513], and enhance explainability and dynamic adaptability [360, 419]. While effective, they are predominantly designed for texts and images and not readily
transferable to graph-structured data for GraphRAG for two reasons. First, the input/output format of
GraphRAG differs significantly from that of traditional RAG. While most retrievers in RAG use NLP
tokenizers for encoders and adhere to the "Text-in, Text-out" workflow, the workflow of GraphRAG
is more diverse, including "Text-in, Text-out" [395, 428], "Text-in, Graph-out" [454, 569], "Graph-in,
Text-out" and "Graph-in, Graph-out" processes [433]. Secondly, retrievers in traditional RAGs do
not capture graph structure signals. Methods like BM25 and TF-IDF [337, 333] primarily focus on
lexical signals, and deep-learning-based retrievers [196] usually capture semantic signals, both of
which overlook the graph structure signals. This motivates us to review existing GraphRAG retrievers,
i.e., heuristic-based, learning-based, and domain-specific retrievers, with a particular emphasis on
their unique technical design adapted to graph-structured data.
**2.4.1** **Heuristic-based Retriever**
Heuristic-based retrievers primarily use predefined rules, domain-specific insights, and hard-coded
algorithms to extract relevant information from graph data sources. Their reliance on explicit rules
often makes them more time/resource-efficient compared to deep learning models. For instance,
simple graph traversal methods like BFS or DFS can be executed in linear time without needing
training data. However, this same reliance on fixed heuristics also limits their adaptability to generalize
to unseen scenarios. In the following, we review the heuristic-based retrievers commonly used in
GraphRAG.
**Entity Linking: In heuristic-based retrievers, entity linking involves mapping entities identified**
in the query to corresponding nodes in graph data sources. This mapping forms an initial bridge
between the query and the graph, serving as either the retriever by itself or as a foundation for further
graph traversal to broaden the scope of the retrieval. The effectiveness of this approach relies on
-----
accurate entity recognition conducted by the query processor and the quality of labeled entities on
graph nodes. This technique is commonly applied in knowledge graphs, where Top-K nodes are
selected as starting points based on their textual similarity to the query. The similarity metric can
be computed using vector embeddings [443, 347] and lexical features [428]. More recently, LLMs
have been used as knowledgeable context augmenters to generate mention-centered descriptions as
additional input to augment the long-tail entities where their limited training data usually cause the
entity linking model to struggle to disambiguate [464].
**Relational Matching: Relational matching, similar to entity linking, is a heuristic-based retrieval**
approach designed to identify edges within graph data sources that align with the relations specified
in a query. This method is crucial for tasks that focus on identifying relationships among entities in
a graph. The matched edges guide the traversal process by indicating which edges to explore next
based on the entities and relations encountered in the graph data sources. Similar to entity linking,
Top-K edges are selected based on their similarity to each edge in the graph [200, 119].
In addition to the efficiency and simplicity of the above two types of heuristic-based retrievers,
another key advantage is their ability to overcome ambiguity. For example, although machine/deep
learning-based retrievers are difficult to differentiate semantically/lexically similar entities/relations
(e.g., Byte vs. Bit, and President of vs. Resident of), these heuristic methods can easily distinguish
them based on pre-defined rules, even in cases where semantic/lexical differences are subtle.
**Graph Traversal: After performing entity linking and relational matching to identify initial nodes**
and relations in graph data sources, graph traversal algorithms (e.g., BFS, DFS) can expand this
set to uncover additional query-relevant information. However, a core challenge for traversalbased retrieval is the risk of information overload, as the exponentially expanding neighborhood
often includes substantial irrelevant content. To address this, current traversal techniques integrate
adaptive retrieval and filtering processes, selectively exploring the most relevant neighboring nodes
and incrementally refining the retrieved content to minimize noise. This graph traversal is mainly
used in GraphRAG for knowledge and document graphs. When traversing on these two types of
graphs, many methods extract all paths less than length l between the nodes identified by entity
linking [492, 493, 530, 185, 110], while others consider the l-hop subgraph around the initial
entities [308, 395, 181, 205]. To more efficiently traverse the KG, other methods prune irrelevant
paths via the use of a LLM [271, 347, 428, 134] and others use pre-defined rules or templates to
traverse the graph [406, 246, 72].
**Graph Kernel: Compared with the above heuristic-based retrievers for retrieving nodes, edges,**
and their combined subgraphs, some earlier works (e.g., graph extraction and image retrieval) [448,
218, 123] treat the text and image as the entire graph and use graph-level heuristics such as graph
kernels to measure similarity and retrieve. Graph kernels measure pairwise similarities by calculating
inner products between graphs, aligning both structural and semantic aspects of the query and the
retrieved graphs. Notable examples include the random-walk kernel and the Weisfeiler Leman
kernel [357, 403]. The random walk kernel computes similarity by performing simultaneous random
walks on two graphs and counting the number of matching paths. The Weisfeiler Leman kernel
iteratively applies the Weisfeiler Leman algorithm to produce color distributions of node labels at
each iteration and then calculates similarity based on the inner products of these histogram vectors.
For example, Wu et al. [448] constructs event graphs of both documents and queries and uses a
product graph kernel that counts walks between two graphs to measure the query-document similarity
and rank the documents. Lebrun et al. [218] conducts event graph matching by introducing a fast and
efficient graph-matching kernel for image retrieval. Similarly, Glavaš and Šnajder [123] translates
images into representative attribute structural graphs that capture spatial relations among regions and
perform graph kernel based on random walks to derive hash codes for image retrieval.
**Domain Expertise: The domain-agnostic nature of traditional heuristic-based methods restricts**
their effectiveness in areas that require specialized expertise. For instance, in drug discovery,
chemists typically design drugs by referencing existing molecules with desirable properties rather
than constructing molecular structures from scratch. These molecules are selected based on domain
knowledge that guides the retrieval of structures with similar characteristics. Following this intuition,
many GraphRAG systems incorporate domain expertise to enhance retriever design. Wang et al.
[434] develop a hybrid retrieval system that integrates heuristic-based and learning-based retrieval to
retrieve exemplar molecules that partially meet the target design criteria.
-----
**2.4.2** **Learning-based Retriever**
One significant limitation of heuristic-based retrievers is their over-reliance on pre-defined rules,
which limits their generalizability to data that does not strictly adhere to these rules. For example,
when confronted with entities that have slight semantic or structural variations, such as "doctor"
and "physician", heuristic-based retrievers like entity linking may treat them differently due to their
distinct lexical representations, despite their shared underlying meaning. To overcome this limitation,
learning-based retrievers have been proposed to capture deeper, more abstract, and task-relevant
relations between the query and objects in data sources, which avoid relying solely on hard-coded
rules. These retrievers often work by uniformly compressing information of various formats (e.g.,
texts and images) into embeddings based on machine learning encoders and then fetching relevant
information by conducting an embedding-based similarity search. Notably, some entity linking and
relational matching methods that employ machine learning encoders to generate embeddings for
matching should also be considered as learning-based retrievers.
In conventional RAG, assuming the query q and data sources that contain n instances S are embedded
by corresponding encoders as q = Fq(q) ∈ R[d] and S = FS(S) ∈ R[n][×][d], we retrieve top-k instances
by similarity search according to the pre-defined similarity function ϕ in the embedding space.
_S_ _[∗]_ = arg max _ϕ(q, S),_ (2)
_k_
Unlike RAGs that use language and vision encoders to embed texts and images, encoders used
in GraphRAG retrieval extend beyond independently and identically distributed (i.i.d.) data by
embedding nodes, edges, and (sub)graphs. Depending on the input format, the encoder could be a
text encoder for query, a graph-based encoder for graph structure, and an integrated text-and-graph
encoder for the textual attributed graph [55, 56]. We specifically focus on graph-based encoders.
Existing graph-based encoders can be broadly categorized into shallow embedding methods – such
as Node2Vec and DeepWalk – and deep embedding methods like Graph Neural Networks (GNNs).
Below, we review these two encoders and their unique roles in GraphRAG.
**Shallow Embedding Methods: Shallow embedding methods [114], like Node2Vec [127] and**
Role2Vec [5], learn node, edge, and graph embeddings that retain the essential structural information
of the original graph. Based on the type of structural information that can be extracted, these methods
generally fall into two categories: proximity/role-based embeddings. Proximity-based methods,
such as DeepWalk and Node2Vec [127, 321], focus on preserving the proximity of connected nodes,
ensuring that nodes close in the graph also remain close in the embedding space. Role-based methods,
like Role2Vec and GraphWave [5, 93], generate node embeddings based on their structural roles
rather than their proximity relations. In general, these methods initialize each node with a latent
embedding vector and conduct unsupervised training to squeeze structural signals derived from graph
structure into the embedding. In GraphRAG, proximity-based shallow embeddings can effectively
retrieve entities that are proximally close, while role-based embeddings can capture entities that share
similar roles. For instance, proximity-based embeddings could be used to retrieve academic papers by
fetching papers sharing similar research topics or retrieve reviews from products that are co-purchased
with the current product [346, 429]. Meanwhile, role-based embeddings could support tasks like
generating company emails by retrieving similar emails based on shared roles or tones [429].
**Deep Embedding Methods: Although shallow embedding methods incorporate structural signals**
into learned embeddings for nodes, edges, or entire graphs, they struggle to leverage semantic
features—like bag-of-words representations for academic paper retrieval or atomic numbers for
molecular retrieval [114]. Additionally, these methods lack inductivity, requiring re-initialization and
retraining whenever new nodes, edges, or graphs are added. This limitation significantly reduces
their applicability in GraphRAG retrieval tasks as real-world knowledge evolves dynamically where
new information continually replaces outdated content, such as in citation networks, social graphs,
and knowledge graphs [59, 419, 453]. To address these limitations, deep embedding methods have
been proposed, which not only jointly fuse features and graph structures to obtain embeddings for
retrieval but also inherent inductive property as the newly coming nodes/edges/graphs share common
feature space with the ones during the training phase. One of the most representative and powerful
approaches in this category is GNN, which combines the power of message-passing to encode
structural signals and feature transformation to extract task-relevant information. Mathematically,
_l[th]-layer graph convolution can be formulated as:_
-----
**x[l]i** [=][ γ][Θ]γ
**e[l]ij** [=][ γ][Θ]γ
x[l]i[−][1] _⊕_ � _ϕΘϕ_ �x[l]i[−][1], xj[l][−][1], eij� _,_ Node-level (3)
_j∈Ni_
**e[l]ij** [=][ γ][Θ]γ e[l]ij[−][1] _⊕_ � _ϕΘϕ_ �e[l]ij[−][1][,][ e]mn[l][−][1][,][ x][e]ij _[∩][e]mn�_ _,_ Edge-level (4)
_emn∈Nij[e]_
**G[l]** = ρΘρG ({x[l]i[,][ e][l]ij _[|][ v][i]_ _[∈V][G][, e][ij]_ _[∈E][G][}][)][,]_ Graph-level (5)
In node-level graph convolution, each node vi adaptively aggregates the embeddings of its neighboring
nodes Ni, with weights based on edge features via the weighting function ϕΘϕ. The aggregated
neighborhood embeddings are then combined with the node’s own embedding from the previous
layer x[l]i[−][1], using a combination function γΘγ, as shown by Eq (3). Optimizing loss from training
downstream tasks would enable the weighting function ϕΘϕ to prioritize the most important neighbors
and enable the combination function γΘγ to balance contributions from the node’s neighborhood
and its own embedding. Similarly, in edge-level graph convolution, the same aggregation principle
applies, but the neighbors of an edge are edges incident to the same ending points of that edge
_Nij[e]_ [, as shown by Eq][ (][4][)][. Graph-level embeddings could be obtained by further applying pooling]
operation ρΘρ over node and edge embeddings, as shown by Eq (5). Following this GNN-based
embedding paradigm, various forms of graph knowledge from diverse sources—such as nodes, edges,
and (sub)graphs—can be uniformly embedded into vector representations, as shown in Figure 5(c)
where we derive embeddings for nodes (X), edges (E), and graphs (G).
Having obtained these node/edge/graph-level embeddings further enables us to create embeddings
for different types of structures (S) by combining these sub-structure embeddings according to
specific configurations for each structure. For instance, if the retrieved subgraph is a path within a
knowledge graph, we can aggregate the embeddings of the nodes and relations along that path to
form a cohesive path embedding.[2] Eventually, the resulting embeddings for different structures can
be utilized either during the training phase to optimize query alignment or during the testing phase
to enable similarity-based neural search. For example, GNN-RAG [347] uses a GNN to perform
retrieval, where a separate round of message passing is performed for each query. The query _Q[ˆ] is_
incorporated into the message passing by, including its embedding in the message computation. A
set of “candidate” nodes is chosen which have a probability of being relevant greater than some
threshold. The shortest path from the query nodes to each candidate node is retrieved as context.
Liu et al. [251] consider the use of a conditional GNN [163] where only the linked entities from the
query are initialized to a non-zero representation. The candidate nodes are chosen in a similar manner
to [347]. A single path is then retrieved for each candidate node and is extracted by backtracking
until we reach a query node. REANO [106] encodes the query information into an edge-specific
attention weight, conditional on the query. After aggregation, the top k triples most similar to the
query are chosen as context.
**2.4.3** **Advanced Retrieval Strategies**
Real-world queries are often complex and encode multi-aspect intentions, possess structure patterns,
and desire multi-hop reasoning that the aforementioned basic retrievers struggle to address. For
example, answering "What is the name of the fight song of the university whose main campus
is in Lawrence, Kansas, and whose branch campuses are in the Kansas City metropolitan area?"
demands multi-hop reasoning to identify the university based on location and retrieve information
about its fight song [145, 399]. Similarly, a query like "What are the main themes in the dataset?"
requires understanding the product community structure, retrieving themes for each community,
and aggregating the identified themes together to summarize the main theme [98]. Furthermore,
when asking "Who is the most impactful research scholar in deep learning?" the answer could vary
depending on multiple aspects [536], such as the number of citations, the volume of published papers,
or the number of co-authors. Accurately addressing such queries requires a deeper understanding of
the underlying data distribution to discern which aspect the query prioritizes. To address these highly
complex queries, advanced retrieval strategies have been proposed, and we review them as follows:
2Incorporating structural signals may be necessary, a consideration to be addressed in future work.
-----
**Integrated Retrieval: Integrated retrieval combines various types of retrievers to capture relevant**
information by balancing their strengths and weaknesses. Typically, integrated retrieval approaches
are categorized according to which individual retrievers are used in combination, with notable
examples including neural-symbolic retrieval [83, 220, 428] and multimodal retrieval [69, 266].
Since the knowledge stored in graph-structured data exists mostly in symbolic format, neural-symbolic
retrieval is a natural choice for the integrated retrieval strategy in GraphRAG. This strategy interleaves
rule-based patterns for retrieving symbolic knowledge with neural-based signals for retrieving more
abstract and deep knowledge. For example, Luo et al. [272], Wen et al. [443] first expands the
neighbors based on the knowledge of the symbolic knowledge graph and then performs path retrieval
using neural matching. In contrast, Mavromatis and Karypis [289] first utilizes GNNs to retrieve seed
entities (neural retrieval) and then extract the shortest paths from seed entities (symbolic retrieval).
Similarly, Tian et al. [395], Yasunaga et al. [492, 493], Wang et al. [427], Luo et al. [272] fetch the
k-hop neighborhood of the entities mentioned in the current question-answering pair and the session
of user-generated items as the answer candidates (symbolic retrieval) and compute attention between
the query and the extracted subgraph to differentiate candidate relevance (neural retrieval).
**Iterative Retrieval: Iterative retrieval is a multistep process where consecutive retrieval operations**
share common dependencies such as causal, resource, and temporal dependency. These dependencies
can be implicitly characterized by the retrieval order in RAG [399, 481] or explicitly modeled as a
graph structure in GraphRAG [145, 454]. Consequently, iterative retrieval is primarily utilized in
GraphRAG to capture these dependencies. For example, KGP [419] alternates between generating the
next piece of evidence for the question and selecting the most promising neighbor. ToG [381] starts
by identifying initial entities and then iteratively expands reasoning paths until enough information is
gathered to answer the question. StructGPT [181] pre-defines graph interfaces and prompts LLMs to
iteratively invoke these interfaces until sufficient information is collected.
**Adaptive Retrieval: While retrieved external knowledge offers benefits, it also introduces risks.**
If the generator already possesses sufficient internal knowledge for a task, the retrieved external
information may be unnecessary or even conflicting [42, 473]. Specifically, when internal knowledge
fully covers the necessary information, retrieval becomes redundant and may introduce contradictions.
To mitigate this, knowledge checking has been proposed in RAG systems [176, 189, 412, 490]. This
approach allows the system to adaptively assess when and how much external information is needed.
By equipping the retriever with this adaptability, RAG can provide more intelligent, flexible, and
context-aware responses, fostering better harmony between internal and external knowledge sources.
One of the adaptive retrievals in GraphRAG is designed by considering different reasoning depths
for different queries, i.e., too few hops of graph traversal might overlook critical reasoning relations,
while too many can introduce unnecessary noise. Guo et al. [134], Wu et al. [455] address this by
training models to predict the required number of hops for a given query and retrieving the relevant
graph content accordingly. No existing works focus on resolving knowledge conflicts in GraphRAG,
and therefore, we leave this discussion to future work.
**2.5** **Organizer**
After retrieving the relevant content C from external graph data sources, which may be in the format
of entities, relations, triplets, paths or subgraphs, the organizer Ω[Organizer] processes this content in
conjunction with the processed query _Q[ˆ]. The aim is to post-process and refine the retrieved content_
to better adapt it for generator consumption, thereby further improving the quality of the downstream
content generation. Formally, the organizer is represented as follows:
_Cˆ = Ω[Organizer]( ˆQ, C)_ (6)
In GraphRAG, the need for fine-grained organization and refinement of retrieved content is driven
by several key reasons. Firstly, when the retrieved contents are subgraphs, their heterogeneous
format of knowledge in terms of node/edge features and graph structures becomes more likely to
include irrelevant and noisy information, which poses significant difficulty for LLM to digest and
thus compromises the generation quality. This raises the desire for graph pruning techniques to
polish the retrieved subgraph and remove task-irrelevant knowledge. Secondly, LLMs have been
widely demonstrated to possess attention biases toward certain positions of relevant information
within the retrieved context [43]. Therefore, the exponentially growing neighbors as the receptive
-----
field enlarges (i.e., the number of hops increases) in the retrieved subgraphs would also exponentially
increase the amount of context length in the prompt and dilute the focus of LLMs on the taskrelevant knowledge [358]. This poses a new requirement for the graph-based reranking mechanism to
prioritize the most important content within the retrieved graph. Thirdly, the retrieved content might
be incomplete in terms of both the semantic content and the structural content, which necessitates
graph augmentation for the enhancement. Finally, the retrieved content is often a graph, which
not only possesses semantic content information but also owns its unique structure. This complex
structural content is not easily consumed by LLMs that are trained by next-token prediction coupled
with linearized prompting, which requires structure-aware verbalization techniques to reorganize. We
will formally review each of the above-motivated organizer techniques in the following sections.
**2.5.1** **Graph Pruning**
In GraphRAG, the retrieved graph can be large and potentially contain a significant amount of noisy
and redundant information. For example, when graph traversal methods are applied in retrieval, the
size of the retrieved subgraph exponentially increases with the number of hops. Large subgraph sizes
not only increase computational costs but can also reduce generation quality due to the inclusion of
noisy information. In contrast, if the number of hops is too small, the retrieved subgraph may be
too small to include crucial knowledge required by tasks. To achieve a better trade-off between the
size of the retrieved subgraph and the amount of its encoded task-relevant information, various graph
pruning methods have been proposed to reduce the size of subgraphs by removing irrelevant nodes
and edges while preserving the essential information.
- Semantic-based pruning: Semantic-based pruning focuses on reducing the graph size by removing
nodes and edge relations that are semantically irrelevant to the query. For example, QA-GNN [492]
prunes irrelevant nodes with low relevance scores by encoding the query context and node labels
using LLMs, followed by a linear projection. GraphQA [389] further removes clusters of nodes with
the lowest relevance to the query. KnowledgeNavigator [133] scores the relations in the retrieved
graph based on the query and prunes irrelevant relations to reduce graph size. Additionally, Gao
et al. [118] partition the retrieved subgraph into smaller subgraphs and then ranks them with only
the top-k smaller subgraphs retained for generations. G-Retriever [147] defines a semantic score
for each retrieved node and edge, then refines the graph by solving the prize-collecting Steiner tree
problem to construct a more compact and relevant subgraph.
- Syntactic-based pruning: Syntactic-based pruning removes irrelevant nodes from a syntactic
perspective. For instance, Su et al. [377] leverages dependency analysis to generate a parsing tree
of the context and then filters the retrieved nodes based on their span distance from the parsing tree.
- Structure-based pruning: Structure-based pruning methods focus on pruning the retrieved graph
based on its structural properties. For example, RoK [431] filters out reasoning paths in the
subgraph by calculating the average PageRank score for each path. Other works, such as Jiang et al.
[181] and He et al. [144], also leverage PageRank to extract the most relevant entities.
- Dynamic pruning: Unlike the aforementioned methods, which typically prune the graph once,
dynamic pruning removes noisy nodes dynamically during training. For example, JointLK [382]
uses attention weights to recursively remove irrelevant nodes at each layer, keeping only a fixed
ratio of nodes. Similarly, DHLK [430] filters out nodes with attention scores below a certain
threshold dynamically during the learning process.
**2.5.2** **Reranker**
The performance of LLMs can be influenced by the position of relevant information within the
context, whether it appears at the beginning, middle, or end [43]. Additionally, LLMs’ generation is
impacted by the order in which in-context knowledge is provided, with later documents contributing
less than earlier ones [172, 256]. While retrieved information is typically ordered by relevance scores
during the retrieval process, these scores are often based on coarse-grained rankings across a large set
of candidates. Enhancing reordering solely among the retrieved information at a fine-grained level, a
process known as re-ranking, is essential to achieve optimal downstream performance. For example,
Li et al. [234] rerank retrieved triples using a pre-trained cross-encoder. Jiang et al. [185] and Liu
et al. [252] employ pre-trained reranker models to rerank retrieved paths. Yu et al. [498] train a GNN
to rerank the retrieved passages. Liao et al. [246] order the paths by the time they occurred, giving
more emphasis to recent paths.
-----
**2.5.3** **Graph Augmentation**
Graph augmentation aims to enrich the retrieved graph to either enhance the content or improve
the robustness of the generator. This process can involve adding supplementary information to the
retrieved graph, sourced from external data or knowledge embedded within LLMs. There are two
main categories of methods:
- Graph Structure Augmentation: Graph structure augmentation methods involve adding new
nodes and edges to the retrieved graph. For instance, GraphQA [389] augments the retrieved subgraph by incorporating noun phrase chunk nodes extracted from the context. Moreover, Yasunaga
et al. [492] and Taunk et al. [389] treat the query as a node, integrating it into the retrieved graph to
create direct connections between the query and relevant information. Tang et al. [388] augment
the graph structure based on pretrained diffusion models.
- Graph Feature Augmentation: Graph feature augmentation methods focus on enriching the
features of the nodes and edges in the graph. Since the original features might be lengthy or sparse,
data augmenters can be employed to summarize or provide additional details for these features. For
example, Once [258] uses LLMs as Content Summarizers, User Profilers, and Personalized Content
Generators in recommendation systems. Similarly, LLM-Rec [276] and KAR [458] apply various
prompting techniques to enrich node features, making them more informative for downstream
tasks.
Additionally, some graph augmentation techniques focus solely on the retrieved graph itself, such as
randomly dropping nodes, edges, or features to improve model robustness. Ding et al. [86] provide a
systematic review of these data augmentation methods.
**2.5.4** **Verbalizing**
Verbalizing refers to converting retrieved triples, paths or graphs into natural language that can be
consumed by LLMs. There are two main approaches to verbalization: linear verbalization and
model-based verbalization.
Linear verbalization methods typically convert graphs into text using predefined rules. The primary
techniques for linear verbalization include:
- Tuple-based: These methods place the different pieces of retrieved information and order
them in a tuple [14, 309]. For example, when performing retrieval on a KG, many methods retrieve a set of facts. A single fact is verbalized in the generation prompt as the tuple
(entity 1, relation 1, entity 2) [308, 395]. For a set of facts, we first sort them in a specific order,
and then verbalize them one at a time as an individual tuple. Each piece of information is typically
separated by line in the prompt. Note that the same logic can be applied to paths, nodes, and so on.
- Template-based: These methods verbalize paths or graphs using predefined templates to generate
more natural text. For example, LLaGA [49] proposes some templates such as Hop-Field Overview
Template to convert graph into sequence. For KGs, several methods [134, 244] convert individual
facts into natural text. For example, Guo et al. [134] convert a fact (entity 1, relation, entity 2) to
text using the template “The {relation} of {entity 1} is/are: {entity 2}”.
Model-based verbalization methods typically use fine-tuned models or LLMs to convert input facts
into coherent and natural language. These methods generally fall into two categories:
- Graph-to-text verbalization: These methods focus on converting retrieved graphs into natural
language while preserving all the information. For instance, Koncel-Kedziorski et al. [211] and
Wang et al. [421] leverage graph transformers to generate text from knowledge graph. Ribeiro et al.
[336] evaluate several pretrained language models for graph-to-text generation, while Wu et al.
[455], and Agarwal et al. [3] fine-tune LLMs to transform graphs into sentences, ensuring a faithful
representation of the graph content in textual form.
- Graph Summarization: In contrast to Graph-to-Text Verbalization, which retains all details,
Graph Summarization methods aim to generate concise summaries based on the retrieved graph
and the query. EFSum [208] proposes two approaches: one directly prompts LLMs to summarize
the retrieved facts and query, while the other fine-tunes LLMs specifically for summarization
tasks. CoTKR [457], on the other hand, alternates between two operations: Reasoning, where it
-----
decomposes the question, generates a reasoning trace, and identifies the specific knowledge needed
for the current step; and Summarization, where it summarizes the relevant knowledge from the
subgraph that retrieved based on the current reasoning trace.
**2.6** **Generator**
The generator aims to produce the desired output for specific tasks based on the query and the
retrieved information. These tasks can range from discrimination tasks (e.g., node/edge/graph
classification) to generation tasks (e.g., KG-based question answering) and graph generation (e.g.,
molecular generation). Due to the uniqueness of different tasks, different generators are often desired.
We categorize generators into three main types: Discriminative-based Generators, which leverage
models like GNNs and Graph Transformers for tasks like classification; LLM-based Generators,
which utilize the capabilities of LLMs to generate answers for text-based tasks; and Graph-based
Generators, which generate new graphs using generative models such as diffusion models. Next, we
provide a detailed illustration of these generators.
**2.6.1** **Discrimination-based Generator**
Discrimination-based generators focus on discriminative and regression tasks, which can typically be
modeled as graph tasks, such as node, edge, or graph classification and regression. Models designed
for graph data, such as GNNs and Graph Transformers, are widely used as discrimination-based
generators. The choice of GNN depends on the graph type and task. For instance, GCN [206],
GraphSAGE [138], and GAT [402] are typically applied to homogeneous graphs, whereas models
like RGCN [350] and HAN [423] are used for heterogeneous graphs, and HGNN [111] and HyperAttention [16] are suitable for hypergraphs. Additionally, graph transformers [296, 354] have gained
popularity for their ability to capture global dependencies. Additionally, different training strategies,
such as (semi-)supervised learning [279] and graph contrastive learning [192, 264], are employed
depending on the specific requirements of the task.
**2.6.2** **LLM-based Generator**
LLMs have demonstrated remarkable capabilities in understanding and generating natural language
across a wide range of tasks. However, LLMs are inherently designed to process sequential data,
while the retrieved information in GraphRAG is typically structured as graphs. Although various
GraphRAG organizers, such as verbalization methods, convert retrieved graph information into text,
these transformations may result in the loss of important graph structure information, which could
be crucial for certain tasks. To take advantage of the ability of LLMs, many research efforts have
been proposed to feed the graph information into LLMs, and we summarize them into the following
categories:
- Verbalizing: Verbalizing aims to convert the retrieved information in GraphRAG into sequences
that can be processed by LLMs. These methods are detailed in Section 2.5.4.
- Embedding-fusion: Embedding-fusion integrates graph embeddings and text embeddings within
LLMs. The graph embeddings can be obtained using GNNs or Graph Transformers[36]. To align
graph embeddings with text embeddings, a domain projector is typically learned to map graph
embeddings to the text embedding space. Embedding fusion can occur at different layers of LLMs.
For example, He et al. [147] feed the projected graph embeddings through the self-attention layers
of LLMs, while Tian et al. [395] prepend the projected graph embeddings with the text tokens. [9]
fuse the text and projected graph embeddings before the prediction layers of LLMs. Additionally,
LLMs can either be fine-tuned along with the domain projector using methods such as LoRA, or
the LLM can remain fixed, training only the graph embedding model and domain projector.
- Positional embedding-fusion: Directly converting the graph into sequences by Verbalization may
lose graph structure information, which can be crucial in some tasks. Positional embedding-fusion
aims to add the position of nodes in the retrieved graph to the LLMs. GIMLET[549], as a unified
graph-text model, employs a generalized position embedding to encode both graph structures
and textual instructions as unified tokens. LINKGPT [148] leverages the pairwise encoding in
LPFormer [361] to encode the pairwise information between two nodes.
-----
**2.6.3** **Graph-based Generator**
In the scientific graph domain, GraphRAG generators often go beyond LLM-based methods due to
the need for accurate structure generation. RetMol [433] is particularly versatile because it can work
with various encoder and decoder architectures, supporting multiple generative models and molecule
representations. For example, generators can be transformer-based or utilize Graph VAE architectures.
Huang et al. [167] highlight the use of a diffusion model, specifically the 3D molecular diffusion
model IRDIF. In the generation process, SE(3)-equivariance is achieved through architectures like
Equivariant Graph Neural Networks (EGNNs) [349], which ensure that the geometric properties of
molecular structures remain invariant to spatial transformations such as rotation, translation, and
reflection at each step. Incorporating SE(3)-equivariance into the diffusion model guarantees that
the generated molecular structures maintain geometric consistency under these transformations. For
KGs, multiple works [110, 492, 389] use a GNN to generate the answer. The GNNs used in these
works are conditional on the query, thereby making the final predictions relevant to it.
**2.7** **Graph Datasources**
We have conducted a comprehensive review of the primary techniques applied in the initial four
model-centric components of GraphRAG—namely, the query processor, retriever, organizer, and
generator. However, even with the best configurations of these components, a GraphRAG system
may still fall short of optimal performance if the underlying graph data sources, from which external
knowledge is retrieved, are not meticulously curated. This also underscores the recent significant
shift in AI research from a model-centric to a data-centric perspective, where enhancing data quality
and relevance becomes equally, if not more, crucial for achieving superior results. Adopting this
data-centric perspective, the following section provides an overview of existing GraphRAG research
on constructing graph data sources from a high-level perspective, with a detailed discussion of
domain-specific graph construction methods reserved for the subsequent domain-specific section.
- Explicit Construction: Explicit construction refers to building graphs based on explicit and
predefined relationships in the data. This method is widely adopted across various domains. For
example, molecule graphs are constructed from the connections between atoms; knowledge graphs
are formed based on explicit relationships between entities; citation graphs are built by linking
papers through citation relationships; and recommendation graphs model interactions between
users and items.
- Implicit Construction: Implicit construction is used when there are no explicit relationships
between nodes, but instead, implicit connections can be derived. For instance, word co-occurrence
in a document can suggest shared semantic information, and feature interaction in Tabular data can
indicate the correlation between features. Graphs can explicitly model these connections, which
might be beneficial to the downstream tasks.
After the graph is constructed, there are also several ways to formally represent graphs.
- Adjacency matrix: The adjacency matrix is one of the most popular ways to denote a graph.
Specifically, the adjacency matrix A ∈ R[||V×|V|] denotes the graph connections among nodes in V,
where |V| is the number of nodes.
- Edge list: The edge list represents each edge in the graph, typically in the form of tuples or triples,
such as (i, j) or (i, r, j), where i and j are nodes, and r is the relation between nodes i and j.
- Adjacency list: The adjacency list is a node-centric representation where each node is associated
with a list of its neighbors. It is typically represented as a dictionary {i : Ni}, where Ni is the
neighbor list of node i.
- Node Sequence: A node sequence transforms a graph into a sequence of nodes in either an
irreversible or reversible manner. Most serialization methods are irreversible and do not allow for
complete recovery of the original graph structure. For example, there are also some serialization
methods that are reversible which can recover the whole graph structure. For example, Zhao et al.
[552] propose serializing graphs using Eulerian paths by first applying Eulerization to the graph.
Besides, if the graph establishes a tree structure, the BFS/DFS can also serialize the graph in a
reversible manner.
- Nature language: With the growing popularity of LLMs for processing text-based information,
various methods have been developed to describe graphs using natural language.
-----
Note that the above-mentioned data structures can only represent basic graphs without support
for complex scenarios such as multi-relational edges or edge attributes. For instance, using an
adjacency matrix to represent a multi-relational attributed graph requires an expanded structure:
**A ∈** R[|V|×|V|×|R|], where R denotes the set of possible relationships. Here, Ai,j,r represents the
weight of the edge connecting node i and node j under relation r.
Selecting an appropriate graph representation is essential for task-specific requirements. For example,
Ge et al. [122] finds that the order of graph descriptions significantly impacts LLMs’ comprehension
of graph structures and their performance across different tasks.
### 3 Knowledge Graph
A knowledge graph is a structured database that connects entities through well-defined relationships.
It can either encompass a broad spectrum of general knowledge, such as the widely recognized
Google Knowledge Graph [53, 261], or delve deeply into specialized domains, like the BioASQ
dataset [400] for biomedical reasoning. The diverse information contained in a knowledge graph
– represented as entities, relationships, paths, and subgraphs – serves as a valuable resource for
enhancing various downstream tasks across different sectors, including question-answering [395,
428, 493], commonsense reasoning [169], fact-checking [201], recommender systems [131], drug
discovery [28], healthcare [37], and fraud detection [287].
**3.1** **Application Tasks**
This section reviews representative applications that GraphRAG on KGs is used for.
- Question-answering: Question-answering (QA) can focus on a single domain or span across
global knowledge. Typically, a query in text format is given, such as "What is the best way to
predict a baby’s eye color?" or "Were there fossil fuels in the ground when humans evolved?" [395]
– the answer can be a sentence generated by a large language model (LLM), a selected text span
from relevant documents, or even a specific choice in a multiple-choice QA scenario. In all these
contexts, GraphRAG leverages knowledge graphs to retrieve relevant information, providing the
necessary context or supporting facts to generate accurate answers.
- Fact-Checking: Fact-checking is to verify the truthfulness of statements by cross-referencing
them with reliable sources of information. GraphRAG enhances this task by querying a knowledge
graph to retrieve relevant facts and relational structures that either support or refute the given claim.
GraphRAG identifies discrepancies or confirmations within the data by mapping the statement onto
the knowledge graph, providing a thorough and evidence-based validation process.
- Knowledge Graph Completion: Knowledge graph completion is the task of predicting new facts
to enhance the comprehensiveness of the graph and infer missing facts [535]. GraphRAG addresses
this task by retrieving structural knowledge around the triplets for inference, supplying essential
structural knowledge, and enhancing the LLM inference.
- Cybersecurity Analysis and Defense: Cybersecurity Analysis and Defense aims to analyze and
respond to vulnerabilities, weaknesses, attack patterns, and threat tactics. With the increasing
complexity and volume of cybersecurity data, GraphRAG has been proposed to provide cybersecurity analysis with more comprehensive insights into potential attack vectors and mitigation
strategies [330].
**3.2** **Knowledge Graph Construction**
We discuss how KGs are typically constructed. For each type of construction technique, we give
examples of common KG databases. How a KG is constructed is important, as it can affect both its
usefulness and function in different downstream tasks. We describe the main techniques below:
- Manual construction: Some KGs are constructed manually via human annotation. WikiData [404]
is a KG that uses crowd-sourced efforts to gather a variety of knowledge. Each entity corresponds
to a page in the Wikipedia encyclopedia. Another KG, the Unified Medical Language System
(UMLS) [25], contains biomedical facts collected from numerous sources.
-----
- Rule-based construction: Many traditional approaches use rule-based techniques for constructing
the graph. This takes the form of custom parsers and manually defined rules used to extract facts
from raw text. Note that these parsers can differ depending on the source of the text. Prominent
examples include ConceptNet [373], which links different words together via assertions, and
Freebase [27] which contains a wide variety of general facts. REANO [106] extracts the entities
and relations from a set of passages using traditional entity recognition (ER) and relation extraction
(RE) methods, respectively. This includes SpaCy [151] for ER and TAGME [112] for RE. To
extract the facts that connect two entities via a relation, they use DocuNet [526].
- LLM-based construction: Recently, work has explored how LLMs can be used to construct
KGs from a set of documents. In such a way, the LLM can automatically extract the entities and
relations and link those together to form facts in the given text. Of note is that no ground-truth
KG exists for these methods. Rather, they simply use a KG as a way to organize and represent a
set of documents. For example, CuriousLLM [487] considers passages in the text as entities and
determines whether two entities should be connected based on their encoded textual similarity. On
the other hand, Cheng et al. [58] uses a manually-defined prompt to convert a piece of text into
a KG. Graph-RAG [98] first divides each document into chunks and then uses an LLM to detect
all the entities in each chunk, including their name, type, and description. To identify the relation
between any two entities, both entities and a description of their relationship are passed to an LLM.
An LLM is then used again to summarize the content of each entity and relation to arrive at their
final title. Lastly, AutoKG [40] uses a combination of LLM embeddings and clustering techniques
to construct a KG from a set of texts.
**3.3** **Retriever**
Real-world facts in KGs can provide grounded information for generative models, enhancing the
reliability of the model output. Given the structured nature of KGs, they are naturally well-suited
for retrieval. The goal is for a given question or query to retrieve either relevant facts [3] or entities
that can help answer that question. Multiple considerations need to be considered during retrieval,
including the type of facts we want to retrieve, the efficiency, and the amount of facts retrieved. In
general, retrieval of KGs has two stages: identifying seed entities and retrieving facts or entities. We
describe both below.
**Identifying seed entities: The first step in retrieving the relevant facts for a given query is to identify**
a set of “seed entities”, which we’ll refer to as Vseed. Seed entities are the initial entities that are
chosen to be highly relevant to the original query. Given such, we expect that triples that contain
any of these entities or are nearby in the graph should provide helpful context. Multiple techniques
exist for identifying the seed entities. Some works [181, 200, 251, 380, 522] assume that we are
given a set of initial entities for each query. However, most works [110, 381, 443, 530, 308, 493]
attempt to extract the entities from the query. One approach is through entity extraction [6], which
uses methods specifically designed for extracting entities from a given text. Most works only extract
entities from the original query. Another common approach is to extract a set of entities that are
semantically similar to the original query [443, 347]. HyKGE [185] first generates a hypothesis
and extracts entities from the original query and the hypothesis. Similarly, in order to reduce the
possibility of hallucination, Guo et al. [134] uses an LLM to generate two similar questions and
retrieves all entities found in the original and generated questions. In a similar vein, RoK [431] first
uses chain of thoughts reasoning to expand the original query, extracting the seed entities from the
expanded query.
**Retrieval Methods: The outcome of the previous step provides us with a set of entities that are**
related in some capacity to the query. These entities are then leveraged to retrieve a set of facts or
entities that can aid us in answering the query. We summarize the core retrieval methods below.
- Traversal-based retriever: These methods traverse the graph and extract paths to aid in answering
a specific question. Given the set of seed entities, Vseed, Yasunaga et al. [492, 493], Zhang et al.
[530] extract all paths up to length two between the entities in Vseed, resulting in a final entity set V .
They further augment V by including all triples that connect any two entities in V . Given V, both
[492, 530] only keep the top k entities by relevance scores. This is calculated by training a separate
model that takes the text embedding of the query and entity as input and outputs how relevant
3Throughout this paper, we will refer to facts as triples or edges, interchangeably
-----
the entity is to the query. For Yasunaga et al. [493], if |V | > 200, they randomly sample 200
entities. Sun et al. [381] use a version of beam-search to explore the KG. Jiang et al. [185], Feng
et al. [110] extract all paths of length k ≤ 2 between seed entities. Alternatively, LARK [62]
retrieves all facts that lie on paths ≤ _k in length starting from the seed entities. Delile et al. [78]_
first extract the shortest paths connecting all seed entities. They further prioritize some entities
over others by considering the recency, importance, and relevance to the query of their associated
text. OREOLM [158] traverse k hops from the seed entities, contextualizing the importance of
each relation and entity to a path via a learnable d-dimensional embedding and it’s LM-encoded
representation. Zhang et al. [522] introduce a trainable retriever that traverses the graph starting
from each seed entity. They also train a model to score each newly visited edge, only keeping a
portion of them. KG-RAG [347] works in a similar manner, scoring each edge by its relevance
and similarity to the query via a dense retriever. They then use an LLM to decide which paths
to explore in the next step. RoG [271] uses instruction tuning to fine-tune an LLM to generate
useful relation paths, which can be retrieved from the KG. KnowledgeNavigator [134] first uses
the query to predict the expected number of hops, hQ needed to traverse in the retrieval stage. It
then traverses hQ hops starting from the seed entities, using an LLM to score and prune irrelevant
nodes. Wu et al. [456] operate in a similar manner; however, they choose which paths to traverse
based solely on the relations. Furthermore, when scoring a path, all relations that lie on that path
are considered when computing the score. RoK [431] considers a different approach, using the
Personalized PageRank (PPR) score to identify useful paths. They further augment these paths by
including the 1-hop neighbors of the seed entities. PullNet [380] assumes that each entity has an
associated set of documents. Given a single seed entity, PullNet, traverses k hops, where in each
iteration it extracts the facts for the newly observed entities. It also extracts any entities that are
contained in documents associated with an entity found in the traversal. Furthermore, for each
entity, only the top N facts are used, which are ranked via similarity to the query. KG-R3 [312] uses
MINERVA [75], a reinforcement learning approach to mining paths between entities, to retrieve a
set of important paths between both entities in the fact. Wang et al. [428] use an LLM to traverse
the graph starting from the seed entities. At each iteration in the traversal, we choose the next
node to visit by prompting an LLM. Specifically, given the information already collected in the
traversal, the LLM is prompted to generate the remaining information needed to correctly answer
the question. The neighboring node that best matches the required information is chosen as the
next node to visit. They further instruction-tune the LLM.
- Subgraph-based retriever: These methods extract a subgraph of size k around each of the
seed entities. Facts that contain one of the seed entities or are nearby, should be highly relevant
to answering the question. Furthermore, they may actually contain the answer itself. Each of
[308, 395, 181, 205] extract either the one or two hop subgraph around each seed entity. The final
set of facts is the union of each individual subgraph. Gao et al. [118] propose to first extract the
subgraph containing the seed entity and potential answers using the method in Sun et al. [379].
This is then partitioned into a set of smaller subgraphs. Then, they design a framework to rank
the subgraphs, keeping the top k subgraphs for generation. For a question-choice pair, MVPTuning [164] considers the triple that contains the highest number of seed and choice entities. They
further augment this by extracting the top k most similar questions in the dataset using BM25 [338],
and extract the triples for each of them.
- Rule-based retriever: These methods use pre-defined rules or templates to extract paths from the
graph. GenTKG [246] considers a temporal KG, where they first extract logical rules from the KG,
and use the top k rules to extract paths in a given time interval for the seed entities. Both [72, 270]
generate queries using SPARQL, which are then used to retrieve import paths. KEQING [406]
decomposes the original query into k sub-queries using using an LLM fine-tuned via LoRA [153].
For each sub-query, they find the most similar question templates, which are predefined. For each
template, they further pre-define a set of logical chains, which are then used to extract matching
paths for the seed entities in the sub-query from the KG.
- GNN-based retriever: GNN-RAG [289] trains a GNN for the retrieval task. A separate round of
message passing is done for each query q, which is incorporated in the message computation along
with the relation and entity representations. The GNN is then trained as in the node classification
task, where the correct answer entity for q has a label of 1 and 0 otherwise. During inference, the
entities with probability above some threshold are treated as candidate answers, and the shortest
path from the seed entity is extracted. Liu et al. [251] use a conditional GNN [163] for retrieval,
where for each query, only the seed entity (they assume there is only one) is initialized to a non-zero
-----
representation based on the LLM-encoded query. They then run L rounds of message passing
where after each layer l only the top-K new edges are kept, resulting in a set of entities Cq[l] [. This]
is determined by a learnable attention weight, which prunes the other edges from the graph. The
final set of candidate entities is the union of candidate entities at each layer l, Cq = ∪[L]l=1[C]q[l] [. It]
is optimized in a similar manner to [289]. For each candidate entity, they retrieve an evidence
chain by backtracking from the entity until it reaches the seed entity, choosing those edges with
the highest attention weight. REANO [106] initializes the entity and relation representations via
the mean-pooled representations of all mentions of that entity/relation in the texts, encoded by T5.
They then run a GNN, which includes an attention weight that considers the relevance of a given
triple to the original question (also encoded by T5). After running the GNN, they retrieve the top
K triples in the KG that are most relevant to the question, where relevance is defined via the dot
product between the triple encoded by the GNN and the question.
- Similarity-based retriever: STaRK [452] considers the vector similarity of the query to each entity.
Each entity embeds both the textual and relational information together. They further consider
multi-vector similarity, where the entities are encoded using multiple vectors. This is done by
chunking the textual and relational information of each entity, with each chunk being embedded
into its own vector. Both REALM [567] and EMERGE [566] extract the entities most similar to
the query. While REALM only retrieves the entities themselves, EMERGE further retrieves the
1-hop subgraph around each entity.
- Relation-based retriever: Kim et al. [200] propose a general framework for reasoning on KGs
using LLMs. They first use an LLM to segment the original query into a set of i ∈ _I sub-sentences,_
where each sub-sentence Si has an associated set of entities Ei. For each sub-sentence, they further
use an LLM to retrieve the top-k most relevant relations Ri,k. Given the set of k relations, for
each sub-sentence, they retrieve all triples that contain a relation in Ri,k and whose entities are
in [�]i∈I _[S][i][. GenTKGQA [][119][] focuses on temporal KG QA. Like [][200][], they retrieve the top-k]_
relations for the query. They then retrieve all facts that contain one of the top k relations and satisfy
the temporal constraints.
- Fusion-based retriever: These techniques consider a combination of different retrieval techniques.
Mindmap [443] considers extracting some paths ≤ _k hops from the seed entity and the 1-hop sub-_
graph of each seed. These two extracted components are combined into one subgraph. DALK [224]
uses a procedure similar to Mindmap, where they extract both paths and the 1-hop subgraph around
each seed entity. However, they argue that this procedure often results in the retrieval of redundant
or unnecessary information. To remove these facts, they use an LLM to rank the retrieved facts
given both the original question and the subgraph. Only the Top-k most relevant facts are kept.
UniOQA [244] considers two branches for retrieving. The first is a translator, which is a fine-tuned
LLM that generates the answer in a CQL format. the second is a searcher that retrieves the 1-hop
subgraph around the seed entities. When determining the answer, answers from the translator
are prioritized over those from the searcher. KG-Rank [483] considers ranking all triples in the
1-hop neighborhood of the seed entities via the similarity of the relation to the query, the similarity
of each triple to the encoded output of a = LLM(q), and an MMR ranking [35] that uses the
similarity score. Only the top-ranked triples are kept. GrapeQA [389] extends [492] by further
including a set of “extra nodes”, which are the common neighbors of the entities retrieved via a
path-based retriever. They further introduce a clustering-based method for pruning entities that
may be irrelevant to the query. SubgraphRAG [232] considers both GNN and textual information.
For the GNN, they consider initializing the node representations using a one-hot encoding to
differentiate between seed entities and others. A GNN is then run for L layers, resulting in the final
representation sv for a node v. To retrieve the relevant triples, they consider first concatenating
the final node representations for each triple (h, r, t) such that zτ = [sh, st]. The probability of
choosing this triple is then given by p(h, r, t) = MLP([zq, zh, zr, zt, zτ ]), where zq, zh, zr, zt are
the encoded textual representations of the query and the triple (h, r, t), respectively. Only the top K
triples are chosen.
- Agent-based retriever: These techniques use LLM agents to retrieve facts from the KG. KnowledGPT [425] defines a set of tools for searching over a KG. Given a query, they generate a piece of
code to search over the KG that considers the seed entities. The code is then executed over the KG
to find the correct answer. KG-Agent [182] focuses on fine-tuning an LLM to generate the SQL
code for retrieving the correct answer. Using a set of tools, they extract a set of paths that contain
the seed entities. KnowAgent [568] first identifies the relevant actions for the query via a planning
module. Using these actions, they then generate a set of paths that are used for generation.
-----
**Other retrievers: KICGPT [439] is concerned with the task of knowledge graph completion, where**
given a partial fact (h, r, ∗), we want to predict the correct entity ˆe. KICGPT retrieves the entities
by first scoring all possible entities using a traditional KG embedding score function. That is, for a
score function f (·) and a partial fact (h, r, ∗), they compute the set of scores {f (h, r, e) ∀e ∈V}.
They use RotatE [383] for the function f (·), a popular approach. Only the top k entities by score
are retrieved. To supplement their knowledge, they also retrieve all triples with (a) the same relation
as the query and (b) all triples that contain the entity h in the query. These are referred to as the
analogous and supplement triple pools, respectively.
**3.4** **Organizer**
In this subsection, we describe how the retrieved knowledge is organized for generation. More
concretely, this is how the information is formatted when given to the generator. Note that not every
method necessarily has an explicit organizer. We summarize the common methods below:
- Tuple-based organizer: These methods consider each piece of retrieved information
as an ordered triple. For example, it would include a triple in the generation
prompt as “(entity 1, relation 1, entity 2)”. Similarly, a path of length m is given by
“(entity 1, relation 1, entity 2, relation 2, · · ·, entity m)”. The entities and relations are usually
represented either as their names or IDs. Each triple or path is usually listed on a separate line. Many works append the retrieved paths to the original query as additional context
[381, 185, 289, 347, 271, 62, 251, 431, 568]. Other works that retrieve facts instead of paths
operate in a similar manner, where instead they append the triples [308, 567, 72, 246, 483,
181, 200, 119]. Some methods [566, 395] consider only including the retrieved entities as
the context. Given a set of facts, KG-R3 [312] first lists all entities and then relations, i.e.,
“(entity 1, entity 2, · · ·, entity m, · · ·, relation 1, · · ·, relation m − 1)”. Delile et al. [78] consider
a KG where each entity has an associated chunk of text. Each text chunk for an entity is considered
as a different piece of information to be included in the context. Both [395, 119] represent each
entity and relation as an embedding, which is the combination of the LLM and GNN embedding.
Liu et al. [251] further include the probability of each path containing the correct answer given by
the GNN model. MVP-Tuning [164] considers combining multiple facts that share the same subject
and relation to remove redundant information. That is, for a subject-relation pair (subject, relation),
they denote the facts for k possible objects as “subject relation {object 1, · · ·, object k}”. KGAgent [182] stores the current KG information and the historical reasoning programs in lists.
- Text organizer: Wu et al. [456] verbalize the retrieved subgraph by passing each triple to the
LLM and prompting it to convert it to a text representation. MindMap Wen et al. [443] uses a
similar procedure for subgraphs, where each is organized as a path before being passed to the LLM.
Some methods use a set of pre-defined templates to verbalize the triples or paths [134, 244, 205].
Wang et al. [406] experiment with verbalizing either via an LLM or pre-defined question templates,
finding that LLM-based verbalizing works better for ChatGPT while template-based works better
for LLaMA [398]. KICGPT [439] uses a combination of data preprocessing and LLM prompting
to convert the triples to text. StaRK [452] uses an LLM to synthesize each entity with its relational
and textual information. Note that they use some pre-defined templates that depend on the specific
task. CoTKR [457] uses an LLM to summarize and then re-write a subgraph of facts for a question
through a “knowledge rewriter”. To train the rewriter, preference alignment is used, which optimizes
the rewriter’s output to match our preferred output. First, k representations of the retrieved subgraph
are produced, with ChatGPT choosing the best and work representations as the most and least
preferred solutions.
- Other organizer: There are some exceptions to the previous classification. KnowledGPT [425]
represents the information in the form of a python class format. They also experiment with
including additional information like the entity description and entity-aspect information.
- Re-Ranking: Some methods also re-rank the information in a specific order. This is done as the
order of information can have a subtle impact on LLM performance. Delile et al. [78] order the
text chunks of each entity based on the impact (measured by # of citations of the parent paper)
and the recency. Dai et al. [72] sort the triples by the relevancy score to the triple. Choudhary
and Reddy [62] attempt to order the paths in a logical matter, such that for a given path, the
subsequent paths build upon it. Yang et al. [483] re-rank the retrieved triples using a task-specific
Cross-Encoder [187]. STaRK [452] considers re-ranking the retrieved entities using an LLM. The
-----
LLM is given the relational and textual information of each, and is asked to give it a score from 0
to 1, which is then used for re-ranking. GenTKG [246] orders the paths by the time they occurred,
further including the time with each. KICGPT [439] ranks all entities using the score of the KG
embedding score function, keeping only the top k entities. KICGPT re-rank the entities using
in-context learning, where they prompt the LLM with examples from the analogy and supplement
pool, as prior knowledge to aid the LLM in how to re-rank the entities.
**3.5** **Generator**
In this section we describe how the retrieved and organized data is used to generate a final response
to the query. We categorize these generators according to the type of methods used to create these
responses.
- LLM-based generator: The vast majority of works use a LLM to generate the response. The input
to the LLM is the original query and retrieved and organized context, formatted using a specific
template. The most commonly used LLMs include ChatGPT [310], Gemini [390], Mistral [180],
Gemma [391], among others. For open-source models where the weights are publicly available,
fine-tuning is sometimes used to modify the weights for a specific task [244, 568]. This is often
done through LoRA [153], which allows for efficient fine-tuning.
- GNN-based generator: Some methods use graph neural networks (GNNs) [207] to conduct the
generation. Yasunaga et al. [492], Taunk et al. [389], Feng et al. [110] extract both the language and
GNN embeddings for each potential answer (i.e., entity) conditional on the query. The probability
of a single entity being the answer is then learnt based on the fusion of the two types of embeddings.
- Other generators: Zhang et al. [530], Yasunaga et al. [493], Hu et al. [158] formulate the
prediction as a masked language modeling (MLM) problem. The goal is to predict the correct value
(i.e., entity) for the masked token which answers the query. To do so, they fine-tune RoBERTa [263]
language model. KG-R3 [312] scores the potential answer entities by performing cross-attention
the representations of the query and each individual entity. PullNet [380] uses GraftNet [379]
to score the different entities. Gao et al. [118] first selects the correct subgraph by computing
the cosine similarity between the query and subgraph representations. For the subgraph with the
highest similarity, it’s fed to GraftNet [379] to select the most probable entity. REANO [106]
passes the encoded triples and their associated text passages to the T5 decoder. The task is framed
as a classification problem, where the goal is to assign the highest probability to the triple with the
correct answer.
**3.6** **Resources and Tools**
In this section, we list common tools and KGs that are used in graph RAG systems. For each, we
give a brief description and a link to the project.
**3.6.1** **Data Resources**
- Freebase [4] [27] is an encyclopedic KG that contains a large variety of general and basic facts.
- ConceptNet [5] [373] is a semantic graph, where the links in the graph are used to describe the
meaning of different words or ideas.
- WikiData [6] [404] is a crowdsourced knowledge base that functions as a structured analog to the
Wikipedia encyclopedia.
**3.6.2** **Tools**
- Graph RAG [7] [98] is an official open-source implementation of the Graph RAG [98] framework.
It can further be installed via the graphrag python package.
[4https://developers.google.com/freebase](https://developers.google.com/freebase)
[5https://conceptnet.io/](https://conceptnet.io/)
[6https://www.wikidata.org/wiki/Wikidata:Main_Page](https://www.wikidata.org/wiki/Wikidata:Main_Page)
[7https://github.com/microsoft/graphrag](https://github.com/microsoft/graphrag)
-----
- LangChain [8] is an open-source framework for using LLMs with various components and applications, including RAG, where using RAG on KGs is supportive.
### 4 Document Graph
A document graph typically models the connections between different documents or various granularity of documents. It is widely observed in real-world scenarios [351, 367, 471], such as hyperlinks
connecting different websites and citations linking one paper to another. Additionally, a document
graph’s relationships between sentences and entities can explicitly capture semantic and syntactic
contextual information. The structural information in documents can serve as a valuable resource
for GraphRAG, aiding LLMs in various tasks. For example, in the retrieval process in RAG, the
document containing the answer may have less apparent connections with the question, such as in
multi-hop question answering [288]. However, we can identify these relevant documents through their
connections to other documents whose context is strongly aligned with the question’s context [90].
In this section, we will systematically review the document graph.
**4.1** **Application Tasks**
Document graphs are beneficial for a wide range of tasks. In this subsection, we will review the tasks
where document graphs can play a significant role. Although document graphs have not yet been
fully integrated with LLMs, they still offer great potential for enhancing the capabilities of LLMs in
various applications.
- Multi-document Summarization(MDS): Multi-document summarization aims to condense the
contents of multiple documents into a cohesive summary. Summarizing an entire corpus can
involve large volumes of text, often exceeding the context window limitations of LLMs. Document
graphs can help compress the corpus by extracting key components and their relationships, proving
highly beneficial for MDS [491, 408, 237, 519, 462, 515, 482, 231, 45]. These graphs also provide
different levels of granularity for summarization through hierarchical clustering [98].
- Text generation: Text generation focuses on producing coherent and meaningful text. While
text-based RAG models have been widely used to generate more reliable text, document graphs
can further enhance this process by retrieving similar documents using graph topology. For
instance, when writing the abstract of a paper, access to its cited papers in the related work section
can significantly improve writing efficiency, as these referenced papers often contain relevant
knowledge and context [429, 313].
- Document Retrieval: Document retrieval aims to find a list of documents relevant to a given query,
which is a key task in Information Retrieval (IR). The exact query terms may not always appear
together in the candidate document; however, by leveraging the connections between documents,
we can retrieve related documents through those documents that are strongly linked to the query.
Thus, it becomes essential to consider document-level relationships in the retrieval process. Several
works [90, 501, 544, 249] have leveraged document graphs with various granularities to improve
document retrieval and ranking. Rather than retrieving the entire document, graphs can also be
used to retrieve specific segments, such as chunks.
- Document Classification: Document classification is a fundamental task in natural language
processing. Traditional methods often focus on the locality of words, limiting their ability to capture
long-distance and non-consecutive word interactions. Moreover, these methods typically focus on
individual documents, overlooking the relationships between them, where connected documents
often exhibit homophily, meaning they are more likely to share similar labels. Building a graph can
help enhance document classification by leveraging both local and global relationships between
words within a single document [538, 259] and by utilizing document-level relationships [518, 461].
- Question Answering: Question answering aims to provide answers to questions based on information from documents and is a fundamental task for RAG. However, documents used for question
answering can be long, and traditional methods often focus on local structures within these documents, neglecting their global structure, which is crucial for long-range understanding. Additionally,
some multi-hop questions require reasoning across multiple documents, necessitating the use of
[8https://python.langchain.com/v0.1/docs/use_cases/graph/constructing/](https://python.langchain.com/v0.1/docs/use_cases/graph/constructing/)
-----
document-level relationships. Humans often consolidate scattered information into structured
knowledge to streamline the reasoning process and make more accurate judgments, in line with
cognitive load theory [384, 243]. Graph-based methods are well-suited for this task by constructing
word-level document graphs [307, 410], utilizing document-level relationships [147, 428] and
leveraging hierarchical interactions [108].
- Relation Extraction: Relation extraction aims to extract semantic relationships between entities in
text, which often requires local, global, syntactic, and semantic dependencies, especially in the case
of document-level relation extraction. Graphs have been proven to be helpful for document-level
relation extraction [409, 302, 344, 63, 558] by capturing these dependencies more effectively.
In addition to the aforementioned tasks, graphs have also been shown to be helpful in other areas,
such as fake news detection [155], coherence assessment [260], and machine translation [471].
**4.2** **Document Graph Construction**
Different tasks may require different types of document graphs, such as document-level, sentencelevel, or word-level graphs. Therefore, the method of obtaining the document graph plays a crucial
role. There are primarily two approaches to constructing a document graph: explicit construction and
implicit construction.
- Explicit Construction. In real-world scenarios, many documents have explicit connections.
Examples include web pages linked through hyperlinks, academic papers citing other works, and
social media posts connected by reposts, comments, and interactions. In these cases, it is natural
to connect the documents to build a document graph, as the connected documents often share
semantic relationships [320]. For instance, Asai et al. [10] constructed a Wikipedia graph based
on hyperlinks between Wikipedia articles. Li et al. [240, 239] follow the same way to leverage
hyperlinks to construct graphs. Yu et al. [498] built a graph using an external knowledge graph
(KG), where each node represents a retrieved passage mapped to entities within the KG, and two
passage nodes are connected if their mapped entities are linked in the KG. These connections have
been leveraged to pretrain large language models (LLMs), enhancing their performance across
various tasks [494, 572, 157].
- Implicit Construction. Despite the presence of explicit connections between documents, different
components within those documents also exhibit important semantic and syntactic relationships.
The structure of these components can be highly beneficial for various tasks in natural language
processing [450, 294]. For example, semantic parsing graphs, such as Abstract Meaning Representation (AMR) graphs, can be leveraged to enhance information extraction [546, 162] and text
summarization [88]. Additionally, syntactic parsing trees are exploited to understand grammatical
structure and resolve ambiguities [533].
The implicit construction of document graphs is highly diverse, adapting to the specific requirements
of different tasks. For instance, nodes in document graphs can represent various granularity, such as
words, entities, sentences, text segments, paragraphs, documents, or topics. In addition, document
graphs often exhibit heterogeneity, where edges connect different types of nodes. Next, we will
detail the construction of different types of edges:
**– Word-word edge: The word-word edges connect words with semantic or syntactic relations or**
dependencies. There are various methods to establish these connections, including word cooccurrence within a sliding window [408, 501, 518, 249, 76], dependency parsing graphs [470,
386, 471, 516, 326, 401] generated by NLP parsing tools [219, 216], Abstract Meaning
Representation [476, 546, 162, 88, 475] and semantic graphs [90, 370, 313] based on different
representations [17, 359]. Additional techniques include coreference resolution [471, 344,
222, 82, 369], word embedding similarity [410, 415] and using large language models (LLMs)
to extract relationships between words [98]. For entities, the construction process is similar to
that for words, though entity extraction methods are often required to first identify entities [303,
6, 317].
**– Word-Sentence edge: Word-sentence edges connect words to sentences based primarily on**
the relationship of belonging. These edges are constructed by linking words to the sentences
they appear in [408, 155, 63], with edge weights often measured using term frequency-inverse
document frequency (TF-IDF). Besides, Ramesh et al. [332] connects the entities with passage
titles via hyperlinks with existing out-of-box entity linkers.
-----
**– Sentence-Sentence edge: Sentence-sentence edges connect sentences based on their semantic**
similarity or relationships. For example, these edges can be constructed through sentence interactions [491, 284, 165], similarities between TF-IDF representations [237, 45], BM25 [297],
sentence embeddings [260, 410, 525], Part of speech (Pos) feature and N-Gram feature [482].
Additionally, Zheng and Kordjamshidi [556] construct a Semantic Role Labeling(SRL) graph
using AllenNLP-SRL model [359]. This allows for connecting sentences within long documents or across different documents, which is particularly useful for LLMs with limited
context length.
**– Sentence-document edge: Sentence-document edges are constructed by linking sentences to**
the specific documents they belong to [392].
**– Document-document edge: Document-document edges connect documents based on their**
similarities. These edges can be constructed when entities in one document are referenced or
shared by another [392, 90], through document clustering [410], embedding similarity [237],
topic similarity [525] or structural similarity [260].
Document graphs typically exhibit heterogeneity, consisting of multiple types of edges. In addition,
several hierarchical graphs have been proposed [444, 540], where different levels of abstraction are
captured. The graph structure can also be dynamic or updated during the learning process [408, 302,
410, 259]. Moreover, the node in the document graph can also be the response of LLMs. For example,
GoR [520] connects the history responses of LLMs with the responding chunks of documents for
long-context summarization.
**4.3** **Retriever**
The retrievers for document graphs typically follow the general retriever design, as described in
Section 2.4. However, there are some special retrieval methods as follows:
- Pre-Retrieval: As introduced earlier, there are many scenarios where building a document graph
is necessary. However, constructing a fine-grained graph for a large volume of documents can
be inefficient and unnecessary. The pre-retrieval aims to first retrieve relevant documents based
on the query and then construct a graph. For example, Thayaparan et al. [392] use pre-trained
GloVe vectors to first extract relevant sentences and then construct a graph based on the retrieved
sentences. Zheng and Kordjamshidi [556], Yu et al. [498] also construct graphs based on the
retrieved information.
- Graph similarity-based retriever: Graph similarity aims to measure the similarity between two
graphs. If both the query and the retrieval data source are graphs, calculating the graph similarity
is essential to retrieve relevant information. For instance, [544] leverages the General Maximum
Common Subgraph (GMCS) method to retrieve related graphs.
- Iterative Retriever: In certain tasks, such as multi-step question answering, the node containing the
answer might not be directly similar to the query. Iterative retrievers address this by first retrieving
nodes related to the query and then using the retrieved information to iteratively retrieve subsequent
nodes. For example, Wang et al. [428], Zhang et al. [541] utilize iterative retrieval methods for
multi-step question-answering tasks, and Ma et al. [278] iteratively retrieve the documents based
on the existing knowledge graph. Asai et al. [10] train a Recurrent Neural Network (RNN) to
recurrently retrieve relevant information with the Bayesian Personalized Ranking (BRR) loss.
- Topology-based Retriever: Various topological relationships in a graph can be used to measure
different types of similarity. For example, proximity-based topological similarity measures the
structural distance between two nodes, while role-based topological similarity assesses the similarity
of the roles nodes play within the graph. These similarities can also be leveraged in the retrieval
process [429].
**4.4** **Organizer**
In this section, we describe the organizers for document graphs. GraphRAG on document graphs is
still in its early stage, and many works do not incorporate explicit organizers, as shown in Section 2.5.
We summarize the existing methods below:
- Graph Pruning: Graph Pruning refines the retrieved subgraph to reduce irrelevant information and
improve computational efficiency. For example, Hemmati and Ghassem-Sani [149] prune graphs
-----
based on the local clustering coefficient, while Zhang et al. [539] employs path-centric pruning to
incorporate off-path information. Li et al. [229] dynamically drop irrelevant nodes during decoding,
and Angelova and Weikum [8] prune edges based on a similarity threshold. Additionally, Edge
et al. [98] uses community detection algorithms to create distinct communities that are then fed
into the generators.
- Reranking: Reranking methods aim to reorder retrieved information to facilitate the generation.
For example, Yu et al. [498], Zhang et al. [541], and Dong et al. [90] use GNNs to rerank retrieved
passages. Li et al. [239] perform listwise reranking to reorder passages expanded via the graph
structure.
**4.5** **Generator**
In this section, we summarize commonly used generators for document graphs. Various methods
are applied within document graphs, depending on the input format. Some approaches use the
entire graph as input, in which case GNNs and Graph Transformers are commonly employed. Other
methods take individual sentences as input, where RNNs or (Large) Language Models are suitable.
Additionally, some works leverage both graph and text as inputs, requiring integrated methods that
can process multimodal data effectively. We detail each category in the following:
- GNN-based generator: Numerous works model tasks as graph-related problems, leveraging
GNNs as generators. Conventional GNNs, such as GCN [206], GraphSAGE [138], and GAT [402],
are widely adopted in various studies [320, 392, 491, 300, 313, 90]. When graphs contain edge
relations, Relational Graph Convolutional Networks (R-GCNs) are often used to capture these
relationships effectively [76, 297]. In addition, some works incorporate graph contrastive learning
techniques to enhance performance [462, 524, 149].
- Graph Transformer-based generator: Graph transformers [504], which capture global information across the graph, are used to encode graph structures for various tasks [518, 292]. These
models leverage the transformer’s self-attention mechanism to capture dependencies beyond local
neighborhoods, making them well-suited for tasks requiring global context.
- RNN-based generator: Recurrent Neural Networks (RNNs), such as LSTMs, are popular for
processing sequences. Therefore, RNNs are employed when the input is in text form [369].
- LLM-based generator: LLM-based generators typically transform the retrieved subgraph into text
before using large language models (LLMs) for generation [428, 98]. Various models are used in
this approach. For instance, BERT [81] has been applied by [401, 233, 541] to support generation
tasks, while Yu et al. [498] and Ju et al. [191] leverage the T5 model [329], and Chen et al. [48]
use RoBERTa [263]
- Integrated Generator: Some works leverage both graph and text data simultaneously for generation, using integrated generators that combine graph models with text generation models. For
example, Xu et al. [476] and Wang et al. [410] employ a combination of RoBERTa and GCN,
while Ramesh et al. [332] fuse GNN with T5 to harness the strengths of both graph structure and
language model capabilities.
In addition to the aforementioned generators, some approaches embed graphs directly into text
generation models. For example, Li et al. [237] proposes replacing traditional self-attention layers
in transformers with graph-informed self-attention, enabling the model to integrate graph structure
directly into the generation process.
**4.6** **Resources and Tools**
The data resources for document GraphRAG can include any type of document, and thus, we do not
provide a comprehensive list here. Instead, in the following, we introduce several tools specifically
designed or commonly used for GraphRAG on document graphs:
- CoreNLP: [9] The Stanford CoreNLP natural language processing toolkit offers a comprehensive set of natural language processing tools, including token and sentence boundaries, parts of
speech, named entities, numeric and time values, dependency and constituency parses, coreference,
[9https://github.com/stanfordnlp/CoreNLP](https://github.com/stanfordnlp/CoreNLP)
-----
sentiment, quote attributions, and relations. These tools are valuable for constructing document
graphs.
- spaCy: [10] The spaCy is an advanced natural language processing library known for its speed
and neural network models, which are optimized for tasks such as tagging, parsing, named entity
recognition, text classification, part-of-speech tagging, dependency parsing, sentence segmentation,
lemmatization, morphological analysis, and entity linking.
- BLINK: [11] BLINK is an entity linking python library that uses Wikipedia as the target knowledge
base.
- OpenIE[12]: Open Information Extraction (OpenIE) is a tool for extracting structured information
from text. It is valuable for generating triples from unstructured text, which can be used as nodes
and edges in document graphs.
- CogComp NLP[13]: Developed by the Cognitive Computation Group, this suite includes tools
for natural language processing tasks such as named entity recognition, sentiment analysis, and
coreference resolution.
- GraphRAG [98][14]: GraphRAG is a data pipeline and transformation suite that is designed to
extract meaningful, structured data from unstructured text using the power of LLMs. The process
involves extracting a knowledge graph out of raw text, building a community hierarchy, generating
summaries for these communities, and then leveraging these structures when performing RAGbased tasks.
- LangChain [15]: LangChain is a framework for developing applications powered by LLMs, with
use cases in document analysis and summarization, RAG, chatbots, and code analysis. Notably,
LangChain supports Graph Transformers, which convert documents into graph-structured formats,
making it highly suitable for processing document graphs.
- Neo4j[16]: Neo4j is a graph database platform offering comprehensive tools for storing, visualizing,
managing, and querying graph data. It includes an LLM Graph Builder, which can extract graphs
with LLMs. It also provides GraphRAG demos to demonstrate how to implement a LLMs and
RAG system with Neo4j.
- LlamaIndex [253][17]: LlamaIndex is a data framework designed to support the development of
applications based on LLMs. It allows developers to seamlessly integrate data sources, ranging
from various file formats to applications and databases, with LLMs. LlamaIndex features an
efficient data retrieval and query interface, enabling developers to input any LLM prompt and
receive context-rich, knowledge-enhanced outputs. Notably, it includes a Property Graph Index
that facilitates the building, modeling, storage, and querying of graphs.
- Haystack[18]: Haystack is an end-to-end framework for building applications powered by LLMs,
Transformer models, vector search, and more. It supports a variety of use cases, including RAG,
document search, question answering, and answer generation. Haystack enables the orchestration
of state-of-the-art embedding models and LLMs into custom pipelines for creating comprehensive
NLP applications. Additionally, it supports Neo4j as a DocumentStore, making it suitable for
document graph storage and query operations.
### 5 Scientific Graph
Scientific graphs refer to graph-structured data used in domains such as drug discovery [433, 167, 265]
and biomedicine [484, 255, 78, 185, 30, 443], both of which are common application areas for
GraphRAG. Therefore, in this section, scientific graphs specifically refer to molecular graphs and
medical graphs.
[10https://github.com/explosion/spaCy](https://github.com/explosion/spaCy)
[11https://github.com/facebookresearch/BLINK](https://github.com/facebookresearch/BLINK)
[12https://github.com/dair-iitd/OpenIE-standalone](https://github.com/dair-iitd/OpenIE-standalone)
[13https://github.com/CogComp/cogcomp-nlp](https://github.com/CogComp/cogcomp-nlp)
[14https://github.com/microsoft/graphrag](https://github.com/microsoft/graphrag)
[15https://www.langchain.com/](https://www.langchain.com/)
[16https://neo4j.com/](https://neo4j.com/)
[17https://www.llamaindex.ai/](https://www.llamaindex.ai/)
[18https://haystack.deepset.ai/](https://haystack.deepset.ai/)
-----
In recent years, significant progress has been made in the development of artificial intelligence
for science [1, 193, 393, 283]. Machine learning (ML) and deep neural network technologies are
increasingly driving scientific discovery from experimental data. Notably, generative models such
as Large Language Models (LLMs) have achieved remarkable success in working with scientific
graph data, including molecular graphs and biomedical graphs. In molecular graphs, for example,
atoms serve as nodes, while chemical bonds represent edges, capturing the structure of molecules. AI
technologies can handle both prediction and generation tasks on these graphs, driving advancements
in fields like drug discovery.
Despite the impressive capabilities of generative models like LLMs, several challenges still hinder
their applications to scientific domains. One of the most prominent challenges is the lack of domainspecific expertise. In fields like drug discovery, molecule generation is crucial, yet traditional
generative models often struggle with producing incorrect or scientifically invalid structures. In
medical question-answering (QA) tasks, incorrect answers, hallucinations, and limited interpretability
are frequently encountered. To address these challenges, recent works have proposed leveraging
external knowledge databases to enhance generation via GraphRAG. GraphRAG improves accuracy
by retrieving relevant scientific graphs from extensive databases to guide the generation or answering
process. This approach ensures scientific validity by incorporating known valid graph structures,
leverages existing knowledge for practical applications, and accelerates the generation process by
narrowing the search space.
**5.1** **Application Tasks**
Scientific graphs are beneficial for a wide range of tasks. In this subsection, we will review some
representative tasks where scientific graphs can play a significant role.
- Molecule generation: Molecule generation refers to the process of creating or designing new
molecular structures, often using generative models [433, 167]. It plays a crucial role in fields
like drug discovery. The use of scientific graphs, especially molecular graphs, can enhance the
rationality and accuracy of generated molecular structures in molecular generation. Typically, an
inquiry such as a molecule is given to retrieve the most relevant molecular structures to guide
molecular generation.
- Molecule property prediction: Molecule property prediction refers to the use of computational
methods to estimate the physical, chemical, or biological properties of molecules based on their
structure [265]. It has proven highly effective in accelerating the drug discovery process while
significantly reducing associated costs. The use of scientific graphs, especially molecular graphs,
can enhance the accuracy of prediction. To achieve this, a query molecule is provided, and similar
molecules are identified as demonstrations to improve the prediction.
- Question answering: Question answering (QA) in the scientific domain refers to the use of
computational methods to provide accurate and context-specific answers to scientific questions [484,
185, 443, 225, 175, 449, 366]. This involves retrieving or generating information from scientific
literature, databases, or other resources to address complex, domain-specific queries, such as "After
meals, I feel a bit of stomach reflux. What medication should I take for it?" In graghRAG, the
scientific literature is usually converted into a knowledge graph to provide a basis for answering
questions or to enrich the query.
**5.2** **Scientific Graph Construction**
In GraphRAG, the choice and construction of data sources are crucial and typically include both
public and private datasets. Public datasets are often derived from widely recognized resources, such
as molecular databases like PubChem [204], ChEMBL [121], and ZINC [170], as well as biomedical
literature and data sources like PubMed [269] and ClinicalTrials [509]. These datasets provide a
broad foundation of information across various domains, offering reliable and authoritative references
for the model. On the other hand, private datasets contain user-specific information, such as medical
records from hospitals or confidential clinical trial data [449]. These datasets are highly confidential
and unique, enabling GraphRAG models to offer personalized and proprietary knowledge support.
For chemistry, molecules can be represented in four main forms: 1D SMILES (Simplified Molecular
Input Line Entry System) [433, 265], 2D molecular graphs [274], 3D molecular graphs with coordinates [167], and text captions describing molecular structures [78]. 1D SMILES is a linear string
-----
generated through depth-first search (DFS) on the molecular graph, following specific rules. 2D
molecular graphs represent atoms as nodes and bonds as edges, visually showing the connectivity of
atoms. 3D molecular graphs incorporate spatial coordinates for each atom, reflecting the molecule’s
structure in three-dimensional space, which is crucial for tasks such as molecular docking and reaction
prediction. Text captions, on the other hand, provide a natural language description of the molecular
structure, which can be used in tasks that involve textual data interpreting chemical structures from
text. Scientific graphs can typically be constructed as follows:
- Text-based construction: Text-based graph construction is the most commonly used method,
which can transform textual scientific knowledge into knowledge graphs. Delile et al. [78] believe
that text captions suffer from issues of information redundancy and imbalance. By constructing text
as a knowledge graph, it is possible to rebalance the retrievable information and reduce redundancy.
Building a knowledge graph typically involves two key steps:
**– Entity Extraction: This step involves identifying and extracting key entities from the text, such**
as domain-specific terms (e.g., chemical compounds, genes).
**– Relationship Extraction: After identifying the entities, the next step is to extract the relationships**
between these entities, determining how they are linked in the given context (e.g., "A is related
to B"). This step forms the structural backbone of the knowledge graph.
- SMILES-based construction: Many chemical databases store data in analytical descriptor formats
such as SMILES [434]. To model graphs, SMILES representations need to be transformed
into graph structures. SMILES-based construction utilizes a library such as RDKit [217] to read
the SMILES notation, create a molecular object, and extract atomic and bonding information
to construct a 2D graph. Each atom in the molecule is represented as a node, and each bond is
represented as an edge between nodes. The atoms carry features such as atom type, degree, charge,
and aromaticity, while the bonds can include bond type (single, double, etc.) and whether they are
part of a ring.
- 3D graph construction: Huang et al. [167] introduce IRDIFF, an interaction-based retrievalaugmented 3D molecular diffusion model designed for target-specific molecule generation. For
pretraining, PMINet is utilized to capture interactive structural context information with binding
affinity signals, leveraging the PDBbind v2016 dataset. This dataset offers 3D protein structures
and a substantial set of experimentally validated protein-ligand complexes. These protein structures
are primarily obtained via techniques like X-ray crystallography, nuclear magnetic resonance
(NMR), or cryo-electron microscopy, containing detailed atomic coordinates, bond angles, and
secondary structure information.
**5.3** **Retriever**
The retriever is responsible for locating relevant information based on the input query. In GraphRAG,
this information typically consists of graph-structured data. The retriever can be roughly classified as
heuristic-based retriever and deep learning-based retriever.
- Heuristic-based retriever: A heuristic-based retriever employs predefined rules, algorithms,
and heuristics to identify and retrieve relevant knowledge from graph-structured data sources.
Heuristic-based retrieval can be categorized into several types as below:
**– Similarity-based retriever: Wang et al. [433] retrieve exemplar molecules based on their similarity**
to the input molecule, using cosine similarity as the metric. MindMap [443] encodes the entities
from the query and the external knowledge graph into dense embeddings using BERT, and then
retrieves the entity set with the highest similarity scores.
**– Matching-based retriever: A matching-based method provides an alternative approach, enabling**
LLMs to generate evidence-based responses by leveraging comprehensive private data. Specifically, Wu et al. [449] identify the most relevant graph through a top-down matching process.
Once the relevant content is identified, the LLM generates an intermediate response with its
assistance, enhancing both the transparency and interpretability of the results.
**– Knowledge graph-based retriever: A knowledge graph-based method can effectively identify**
and retrieve the necessary information to respond to a query. Pelletier et al. [318], Li et al.
[225] use named entity recognition and relation extraction to connect user queries with relevant
entities in the knowledge graph, thereby uncovering interpretable and actionable insights from
-----
existing biomedical knowledge. This approach significantly enhances the transparency and
utility of predictive models. Delile et al. [78] map the text chunks to the knowledge graph, then
utilize graph distances to find the chunks most relevant to the user’s question. In addition, this
work introduces a scoring metric that balances the data by giving each concept mapped along
the shortest path an equal opportunity. This metric prioritizes text chunks based on both their
recency and their impact. HyKGE [185] first queries the LLM to generate a hypothetical output
and extracts entities from both the output and the query. Then, HyKGE retrieves reasoning
chains between any two anchor entities in an existing knowledge graph, such as CMeKG [30]
and CPubMed-KG, and feeds the reasoning chains along with the query into the LLM. KGRank [484] identifies entities within the query and retrieves the related triples or sub-graphs from
the KG to gather factual information.
**– Fusion-based retriever: Soman et al. [366] begins by retrieving the relevant node in the knowl-**
edge graph based on vector similarity to the query’s entity. Then, it retrieves the context triples
(Subject, Predicate, Object) linked to this node within the knowledge graph.
- Deep Learning-based Retriever: Deep learning-based retrievers can extract relevant knowledge to
guide the generation process, such as the generation of molecules or proteins. Specifically, Huang
et al. [167] uses a pre-trained protein-molecule interaction network named PMINet to extract
interactive structural context information between the target protein and ligands in the reference
pool to guide the generation of target-aware ligands. Jeong et al. [175] utilize the off-the-shelf
MedCPT retriever, a tool specifically tailored for retrieving documents in response to biomedical
queries, capable of retrieving up to ten relevant pieces of evidence for each input. DALK [225]
filters out noise and retrieves the most relevant knowledge by utilizing the ranking capabilities of
LLMs.
**5.4** **Organizer**
Basically, there are two types of the organizer: embedding-based and query-based organizers according to how the retrieved knowledge is utilized.
- Query-based organizer. The simplest and most direct organizer is a query-based organizer,
which integrates the retrieved information with input queries to generate responses. Specifically,
HyKGE [185] enhances the reasoning process by incorporating external knowledge into the large
language model. In this framework, the retrieved reasoning chains are fed into the LLM alongside
the original query. This enables the model to ground its responses in structured, contextually
relevant information, improving both the accuracy and depth of the generated output. Wu et al.
[449] prompt the LLM to answer the question by providing the retrieved entity names and their
relationships in a concatenated form. This approach involves retrieving relevant entities and the
relationships between them from an external knowledge source and structuring this information
as a cohesive input for the LLM. DALK [225] integrates the query and retrieves knowledge and
then feeds it into LLMs for reasoning and getting the predicted answer. Similarly, KG-Rank [484]
combines the re-ranked triplets with the task prompt and inputs them into LLMs for answer
generation. In contrast to the methods mentioned above, Soman et al. [366] begin with context
pruning. Specifically, this approach refines the retrieved context by selecting only the most
semantically relevant elements needed to answer the query promptly. The input prompt is then
combined with this pruning-aware information, producing an enriched prompt that is fed into
the LLM for text generation. Delile et al. [78] develop a data rebalancing mechanism to ensure
that each entity relevant to a question has an equal opportunity of being represented while also
highlighting recent significant discoveries. This rebalanced knowledge is then combined with the
query prompt and provided as input to the LLM.
- Embeding-based organizer. Embedding-based organizer integrates the retrieved information with
input embedding to generate responses. Specifically, Wang et al. [433] uses a lightweight, trainable
standard cross-attention mechanism to fuse the embedding of the input molecule and the retrieved
example molecules. Similarly, Huang et al. [167] uses a trainable cross-attention mechanism to
fuse the enhanced embeddings of the retrieved example ligands and the generated molecules.
**5.5** **Generator**
The generator is a core component of the GraphRAG model and is responsible for producing the final
output by integrating retrieved evidence with the input data. Generators can be broadly classified into
-----
three categories according to the type of generated models they use: transformer-based generators,
diffusion model-based generators, and large language models-based generators.
- Transformer-based generator: RetMol [433] utilizes the Megatron version of the molecule
generative model Chemformer for drug discovery. Specifically, the Chemformer model is a
Transformer-based model that can be efficiently applied to tasks in chemistry.
- Diffusion-based generator: Huang et al. [167] introduce a novel interaction-based retrievalaugmented diffusion model (IRDIFF) for structure-based drug design. Specifically, IRDIFF is
able to generate molecules that bind strongly to the target pocket by utilizing protein-molecule
interaction data between reference proteins and the target protein to guide the diffusion model.
- LLM-based generator: Some approaches leverage large language models, such as LLaMA2,
LLaMA3, GPT-4, and Gemini, etc [449, 175, 318, 265, 225, 366]. For instance, MedGraphRAG [449] utilizes GraphRAG to improve the answering capabilities of models such as
LLaMA2, LLaMA3, Gemini, and GPT-4 for medical question answering. MolecularGPT [265]
utilizes GraphRAG to improve GPT-3’s predictive capabilities for molecular property prediction.
DALK [225] employs GraphRAG to boost GPT-3.5-turbo’s ability to answer questions related to
Alzheimer’s Disease. HyKGE [185] utilizes GraphRAG to enhance the medical question-answering
capabilities of GPT-3.5 and Baichuan13B.
**5.6** **Resources and Tools**
In this section, we provide an overview of common data sources and tools utilized in graph RAG
systems within the scientific domain, along with a brief description of each project.
**5.6.1** **Data Resources**
The public datasets are commonly from well-known resources, such as molecular databases:
- PubChem [204]: PubChem encompasses a wide range of chemical data, including 2D and 3D
structures, chemical and physical properties, bioactivity, pharmacology, toxicology, drug targets,
metabolism, safety guidelines, associated patents, and scientific literature. Most of its entries
pertain to small molecules, with a primary emphasis on those containing fewer than 100 atoms and
1,000 bonds.
- ChEMBL [121]: It is an openly accessible database containing detailed information on drugs, druglike small molecules, and their bioactivity. This curated resource stands out for its comprehensive
coverage of the drug discovery process, encompassing data on more than 2.2 million compounds
and over 18 million records documenting their effects on biological systems. ChEMBL provides
insights into the interactions between small molecules and their protein targets, along with data on
how these compounds influence cellular and organismal functions. It also includes information on
ADMET (absorption, distribution, metabolism, excretion, and toxicity) profiles. The database stores
two-dimensional structures, calculated molecular properties (such as logP, molecular weight, and
Lipinski’s Rule of Five parameters), and bioactivity data like binding affinities and pharmacological
effects.
- ZINC [170]: The ZINC dataset is a curated collection of commercially available chemical compounds designed specifically for virtual screening purposes. It offers over 230 million purchasable
compounds in 3D formats that are ready for docking, as well as more than 750 million compounds
available for analog searches. Each molecule is adjusted to biologically relevant protonation states
and is annotated with properties such as molecular weight, calculated LogP, and rotatable bonds.
The library includes vendor and purchasing details, making it compatible with several widely used
docking software programs. Compounds are provided in multiple protonation states and tautomeric
forms within certain constraints, and some formats even offer multiple conformations per molecule.
The ZINC database is available for free download [19] in multiple common file formats, including
SMILES, mol2, 3D SDF, and DOCK flexibase.
There are also biomedical data sources as below:
[19http://zinc.docking.org](http://zinc.docking.org)
-----
- PubMed [269]: PubMed is a freely accessible database that primarily houses the MEDLINE
collection, containing references and abstracts in life sciences and biomedicine. Managed by
the United States National Library of Medicine (NLM) within the National Institutes of Health,
PubMed is part of the Entrez retrieval system. As of May 23, 2023, PubMed contains over 35
million citations and abstracts, with records dating back to 1966 and selectively to 1865, and a few
even to 1809. On this date, approximately 24.6 million records include abstracts, and 26.8 million
provide links to full-text articles, with around 10.9 million available freely. In the last decade (up
to December 31, 2019), PubMed added nearly one million new records annually on average.
- ClinicalTrials [509]: ClinicalTrials is a global registry and database that provides information
on clinical studies funded by both private and public sources. Managed by the U.S. National
Library of Medicine, this resource includes a summary of each study’s protocol, and for some
studies, results are available in a tabular format. Users can search studies by criteria such as study
status, condition or disease, country, and other keywords. Continuously updated, the database
now includes over 300,000 research studies conducted across all U.S. states and in more than 200
countries worldwide.
- Open-source medical KGs: CMeKG (Clinical Medicine Knowledge Graph) [20], CPubMed-KG
(Large-scale Chinese Open Medical Knowledge Graph) [21] and Disease-KG (Chinese disease
Knowledge Graph) [22] are open-source medical knowledge graphs that consolidate a vast amount
of medical text data, covering areas such as diseases, medications, symptoms, and diagnostic
treatments. The combined knowledge graph includes 1,288,721 entities and 3,569,427 relations.
**5.6.2** **Tools**
- RDKit [217]: RDKit is open-source toolkit for cheminformatics. RDKit has several key features: it
can process chemical structures by reading and writing various file formats, such as SMILES, InChI,
and Mol files. It generates multiple types of molecular fingerprints, enabling chemical structure
comparison and similarity searches. RDKit also provides algorithms for calculating molecular
similarity and supports the representation and processing of chemical reactions, including the
identification of reactants and products. Although it does not offer molecular docking capabilities
on its own, RDKit can be integrated with other docking tools. Additionally, RDKit supports
machine learning algorithms, allowing for pattern recognition in chemical data and the construction
of predictive models.
- CADRO: The Common Alzheimer’s and Related Dementias Research Ontology (CADRO) [23]
is employed to extract a subset of Alzheimer’s disease (AD)-related samples from the medical
QA datasets for evaluation. CADRO organizes terms into a three-tiered classification system with
eight main categories and multiple subcategories focused on AD and related dementias, containing
frequently used terminologies or keywords in the field. From CADRO, users obtain a list of
AD-related keywords most relevant to the medical QA datasets.
### 6 Social Graph
The social graph typically consists of entities connected by their social relations and is ubiquitous across real-world applications. A primary example is social networks like Twitter and
Facebook, where entities represent individuals linked by social interactions (e.g., friendships,
followers/followees, likes, and mentions). These social graphs extend beyond human interactions and are not confined to living entities, such as tortoises co-using the same burrow in animal social networks [340], complementary products co-purchased by the same customer in Ecommerce recommender systems [427, 452], or even the LLM-simulated social agents [523, 241].
The wealth of social-relational knowledge in these social graphs is the golden resource for
GraphRAG[511, 183, 463, 429, 510, 427, 77, 95, 452, 159, 528, 327, 438, 202, 396], which is
reviewed in this section.
[20https://cmekg.pcl.ac.cn/](https://cmekg.pcl.ac.cn/)
[21https://cpubmed.openi.org.cn/graph/wiki](https://cpubmed.openi.org.cn/graph/wiki)
[22https://github.com/nuolade/disease-kb](https://github.com/nuolade/disease-kb)
[23https://iadrp.nia.nih.gov/about/cadro](https://iadrp.nia.nih.gov/about/cadro)
-----
**6.1** **Application Tasks**
- Entity Property Prediction: Entity property prediction focuses on predicting properties and
classifying categories for social entities in social networks, examples of which include the prediction
of partnership compatibility, assessment of morality, detection of account suspensions, identification
of toxic behaviors [183] and product property prediction [427].
- Text Generation: Text generation aims to produce text that aligns with social contexts and
norms. Typically, the interplay between structural and textual information, such as proximity-based
network homophily and role-based similarity [5], serves as the foundation for text generation.
For instance, Wang et al. [429] retrieves the texts of proximity/role-similar nodes to enhance
the text generation of the target node. Kim et al. [202], Xie et al. [463] generate personalized
recommendation explanations by retrieving customers’/products’ historical reviews. In addition,
some other works also leverage semantic similarity to retrieve reference/attributes/opinions and
augment the downstream review generation [353, 92, 315].
- Recommendation: The recommendation task aims to find the most relevant items to satisfy
user demands. Due to the missing and sparse customer/item interactions (e.g., cold-start issues), GraphRAG can be naturally applied to boost customer/item sparse interaction by retrieving additional meta-knowledge. Depending on the concrete recommendation scenario,
GraphRAG has been used for graph-based recommendation [438, 95], next-item recommendation [427, 159, 528, 327, 510, 416], and conversational recommendation [116]
- Question-answering: Question-answering tasks are encountered not only in knowledge and
document graph domains but also in social graphs. For example, a user might ask, "What are the best
parks for family gatherings around Los Gatos?" and expect a personalized query answer [511, 198].
Furthermore, such questions might explicitly require graph-structured reasoning. For instance, Wu
et al. [452] build a semi-structured knowledge base, and some queries might ask, "Can you list the
products made by Nike?" Answering this question requires not only a deep understanding of the
query but also familiarity with the structural information of the data.
- Fake News Detection: Detecting fake news necessitates considering both semantic content (e.g.,
the content of the news) and structural interactions (e.g., interactions among news). For instance,
Ram et al. [331] evaluates the credibility of a Reddit post based on the credibility of other Reddit
posts retrieved by their common interactions with the same common commenters.
**6.2** **Social Graph Construction**
The relations that GraphRAG leverages to derive additional information from social graphs can be
mainly summarized into three rationales: proximity-based, role-based, and personalization-based
rationale. The proximity-based rationale, rooted in the adage "birds of a feather flock together,"
suggests that nodes close to each other in a social network often share similar properties [291]. For
example, close friends tend to possess similar hobbies. The role-based rationale focuses on nodes
with similar local subgraph structures sharing similar features or label distributions [93]. For example,
managers at the same hierarchical level within a company generally have comparable job titles and
responsibilities, and hub airports exhibit similar operational characteristics and strategic importance.
Lastly, the personalization-based rationale refers to individuals’ uniqueness regarding their characteristics and interactions. For example, in recommender systems, users interact with items in
various ways, such as clicking, viewing, adding to a cart, purchasing, and reviewing, each interaction
providing valuable knowledge that GraphRAG can leverage to personalize generated content. Based
on the above three relational rationales in the social graphs, the social graph construction methods
can be summarized as follows:
- User-User-Interaction [183]: This type of social graph represents user-to-user interactions commonly seen on social networks such as Twitter, Reddit, and Facebook. Examples include followerfollowee relationships on Twitter, friendship relationships on Facebook, and user comments on
other users’ threads on Reddit. Note that this user-user interaction naturally exists in the real world
without human curation or modification compared to documents or knowledge graphs.
- User-Item-Interaction [427, 429, 463, 202]: This type of social graph represents user-to-item
interactions commonly found on e-commerce platforms like Amazon and eBay. These interactions,
including purchasing, adding-to-the-cart, and viewing, can be modeled as a bipartite graph, where
each type of interaction reflects a unique user intention.
-----
- Item-Item-Interaction [427, 528, 452]: This social graph captures interactions between items,
typically identified by shared interactions from the same customer or user. For example, a "co-view"
interaction between two products indicates that they are viewed by the same customer, while a
"view-add-to-cart" interaction indicates that one product is first viewed and the other is added to
the cart next by the same customer. In e-commerce networks, these co-interactions between two
items can be broadly categorized into complementary and substitute relationships [554].
- Metadata-Interaction [452, 485]: Items and users often possess metadata. For example, products
on platforms like Amazon may include brand, manufacturer, and color attributes. This metadata can
be represented as additional node types and the corresponding edges indicating relations between
products and attributes, such as ownership or association.
- Agent-Agent Interaction [51, 523]: With the increasing intelligence of LLMs, recent literature
has explored the potential of using LLM-powered agents to simulate social behaviors, such
as collaboration, debate, and reflection. These agent-agent interactions could also be used for
GraphRAG.
Note that among the five types of interaction mentioned above, the user-user, user-item, and metadata
relations are naturally formulated without human curation. In contrast, item-item interactions are
generated by manual extraction, and agent-agent interactions require simulations.
**6.3** **Retriever**
The relations of the social graphs constructed according to the above methods can be leveraged
in GraphRAG to enhance downstream tasks by retrieving additional information. For example,
retrieving historical metadata and customer interactions can improve recommendations. Next, we
review representative retrieval methods in GraphRAG for social graphs.
- ID-based Retriever: Similar to entity linking, the ID-based retriever works by retrieving content
specifically generated by a user/item, examples of which include retrieving historical item interactions of a specific customer [511, 183, 510], reviews posted on a specific product [463], and
meta-data information about a specific user [452, 346].
- Filtering-based Retriever: Based on the ID-based retriever, the Filtering-based Retriever retrieves
additional content based on collaborative filtering [416]. For user filtering, Top K users are
identified by comparing the similarity of their historical item interactions with those of the target
user demanding recommendation. To enhance the context of the target user, it retrieves the most
popular items from those Top K users. In contrast, for item filtering, it identifies the Top K items
that are most similar to the current item by comparing their user-interaction history. Among them,
the most popular ones would be fetched to augment the current item recommendation.
- Social Relational Retriever: Like the ID-based Retriever, the Social Relational Retriever focuses
on retrieving knowledge from entities sharing certain relations with the target entities on hand.
For example, Du et al. [95] hierarchically retrieve neighboring texts several hops from the central
node to augment the semantic information of the target user/item. Meanwhile, Wang et al. [429]
retrieve texts corresponding to nodes that share high proximity-based and role-based similarity,
respectively.
- Integrated Neural-Symbolic Retriever: This approach leverages both symbolic and neural
retrievers to improve retrieval effectiveness [427, 510, 409]. The symbolic retriever retrieves
information by following explicitly defined rules, such as retrieving based on identifiers, structured
relationships, or interaction patterns, ensuring that the retrieved data strictly aligns with specific
criteria. Meanwhile, the neural retriever complements this by using embedding-based similarity,
capturing nuanced patterns and contextual relationships that may not be directly encoded in rules.
Integrating them together provides a better trade-off between rule-based precision and neural-based
adaptability and generalizability. For instance, Wang et al. [409, 427], Huang et al. [159], Qiu
et al. [327] first retrieve K-hop neighboring products from the product knowledge graph (symbolic
retriever) for products in the user session and further utilize neural-based adaptive filtering to
aggregate items that are most relevant to the current sequence (i.e., neural retriever). Similarly,
Zeng et al. [510] address data sparsity and heterogeneity by combining ID-based retrieval that
retrieves movies based on the user ID of one party with a text-based retriever that enables movie
retrieval between parties.
-----
**6.4** **Organizer**
To further enhance the retrieved content, the organizer for the social graphs employs specialized
techniques beyond the typical re-ranking and filtering used in other graph domains [511, 152, 77].
For social graphs, the organizer of GraphRAG often uses Keyword Extraction, Profile Summarization,
and Hierarchical Graph Aggregation to curate the retrieved content:
- Keyword Extraction: Keyword extraction identifies the most relevant and informative keywords
from the retrieved content. These extracted keywords guide the downstream generator in prioritizing
attention and reduce the risk of overwhelming LLMs with excessive context. For example, Xie
et al. [463] use an embedding estimator to pinpoint keywords aligned with a personalized latent
query, which are then used to generate explanations.
- Profile Summarization: Profile summarization creates rich and detailed user profiles that capture
attributes such as age, gender, preferred and disliked genres, favorite directors, country, and
language derived from users’ past interactions and item metadata [438]. This enriched user data
increases the informativeness of the retrieved content while safeguarding privacy by using rewritten
profiles. Additionally, after retrieving related movies via user/item filtering, Wang and Lim [416]
apply a three-step prompting strategy to extract features tailored to the user and select representative
movies by directly instructing LLMs. These extracted features and representative movies are further
used as the user profile to guide the final recommendation generation. Guo et al. [137] leverage
LLMs to generate more details about the entity profile based on its keywords/short phrases from
external data to aid text generation.
- Hierarchical Graph Aggregation and Summarization: When retrieving neighborhood textual
information, the exponentially expanding receptive field at higher neighborhood layers often
causes the aggregated content to exceed manageable limits, potentially diminishing the LLM’s
effectiveness [236]. This challenge highlights the need for hierarchical graph aggregation and
summarization. The primary idea is to first summarize neighborhood information retrieved at each
layer before propagating it to the next layer. This approach keeps the volume of text each node
receives within consistent bounds. For instance, Du et al. [95] recursively aggregate information
from neighboring nodes at higher layers and leverage LLMs to rephrase and compress the content
before sharing it with relevant nodes in lower layers. Consequently, this strategy expands the
receptive field while optimizing the computational budget for downstream generation tasks.
**6.5** **Generator**
Once the relevant content is retrieved and organized appropriately, it is further processed by a
downstream generator to produce the final content. Depending on the desired output, existing
generators used for social graphs can be categorized into LLM-based text generators and Predictionbased generators. The choice between these two depends on the format of the desired output and the
requirements of the social graph applications.
- LLM-based Text Generation: This generator is typically used when the downstream application
requires text outputs [510, 463, 416, 429], such as item recommendations based on names, recommendation explanations, and review generation. Due to the probabilistic nature of the next token
prediction, the generated text may not always precisely align with the desired output. To address
this hallucination issue, the ground-truth text is often used for grounding the generated texts, such
as matching generated items with the ground-truth items on the platform [152].
- Prediction-based Generation: The generator directly predicts outputs and is primarily used for
non-textual tasks [95], such as item recommendation and social prediction. For instance, Graph
Neural Networks have become popular choices for graph-based recommendation [95, 438] and
transformers are used for item recommendation tasks [427]. Furthermore, in social property
prediction [183], the generator could simply be a multilayer perception used for classification (e.g.,
partisanship classification) or regression (e.g., morality regression).
**6.6** **Resources and Tools**
This section lists common resources and tools used for GraphRAG on social graphs. The summary
and the link of are itemized as follows:
-----
**6.6.1** **Data Resources**
- STARK-Amazon[24] [452]: STARK-Amazon is a large-scale, semi-structured retrieval benchmark
dataset for product search on the Amazon platform, integrating textual and relational knowledge
bases. The nodes in the dataset represent products, colors, brands, and categories, while edges
capture relationships like also_bought, also_viewed, has_brand, has_category, and has_color. Rich
textual information, including product descriptions, customer reviews, and entity names, provides
valuable context for retrieval tasks. Queries are generated by sampling relational templates and
grounding them with specific entities, then leveraging LLMs to synthesize relevant textual and
relational information. This process results in queries that capture customer interests, interpret
specialized descriptions, and deduce relationships involving multiple entities within the query.
- Amazon-Review[25] [306, 146, 290, 463, 510]: The Amazon-Review dataset includes reviews with
ratings, text, and helpfulness votes, along with product metadata such as descriptions, categories,
price, brand, and image features. Additionally, it provides link information through “also viewed”
and “also bought” graphs. This dataset has two versions: the initial release logging the review data
between 1996 and 2014 [146, 290], and a more recent version that logs ongoing data in 2014 [306].
The latest version also adds new metadata, including detailed product information, bullet points,
and an ethical details table. Extensive GraphRAG work [] in social graphs has leveraged this dataset
to build RAG frameworks for recommendation research.
- MovieLens[26] [416, 438, 510]: The MovieLens datasets describe people’s expressed preferences
for movies [143]. These preferences take the form of tuples, each resulting in a person expressing
a preference (a 0-5 star rating) for a movie at a particular time. These preferences were entered
through the MovieLens website — a recommender system that asks its users to give movie ratings
to receive personalized movie recommendations. Side information includes movie title, year, and
genre in textual format. [438] further crawl the visual content of movie posters accessible from
[here. Note that Movielen-100k is frequently used [438, 510] and provides users with demographic](https://github.com/HKUDS/LLMRec)
information such as user ID, age, gender, occupation, and zip code.
- Netflix[27] [438]: This dataset was constructed to support participants in the Netflix Prize. The movie
rating files contain over 100 million ratings from 480 thousand randomly chosen, anonymous
Netflix customers over 17 thousand movie titles. The data were collected between October 1998
and December 2005 and reflect the distribution of all ratings received during this period. The
date of each rating, the title, and the year of release for each movie ID are also provided. The
[multi-model side information is collected through web crawling [438], which is further stored here.](https://github.com/HKUDS/LLMRec)
- Yelp[28] [463]: The Yelp dataset is a rich resource for academic research, particularly in fields
like data science, natural language processing (NLP), and machine learning. This comprehensive
dataset includes over 6.9 million reviews, 150,346 businesses, and 200,100 photos, covering 11
metropolitan areas. Researchers and students can explore real-world business and user data, with
extensive details about businesses such as location, categories, attributes (like hours, parking,
and ambiance), and aggregated check-ins. Reviews include full-text content, user ratings, and
metadata, while user profiles offer insights into social interactions and behaviors, including friends,
compliments, and review history. Additionally, the dataset contains 908,915 tips, providing quick
recommendations and insights, and 1.2 million business attributes, offering granular information
about service offerings.
- Weibo[29] [521]: The Weibo dataset offers a rich snapshot of user interactions, behaviors and social
connections within the Sina Weibo platform, capturing both static and dynamic facets of the social
network. Starting with 100 randomly selected seed users, the dataset scales to encompass 1.7
million users and around 0.4 billion following relationships, averaging 200 followers per user.
Each user is characterized by social profile details, including name, gender, verification status, and
follower/followee counts. Tweet content is provided in both original Chinese and indexed formats,
supporting robust research in social network analysis, content diffusion, and user interaction
dynamics, making it an invaluable resource for RAG studies.
[24https://stark.stanford.edu/dataset_amazon.html](https://stark.stanford.edu/dataset_amazon.html)
[25https://jmcauley.ucsd.edu/data/amazon/](https://jmcauley.ucsd.edu/data/amazon/)
[26https://grouplens.org/datasets/movielens/](https://grouplens.org/datasets/movielens/)
[27https://www.kaggle.com/datasets/netflix-inc/netflix-prize-data](https://www.kaggle.com/datasets/netflix-inc/netflix-prize-data)
[28https://www.yelp.com/dataset](https://www.yelp.com/dataset)
[29https://www.aminer.cn/influencelocality](https://www.aminer.cn/influencelocality)
-----
- Brexit[30] [565]: This dataset includes a portion of the X (Twitter) network, specifically the remainleave discourse before the 2016 UK Referendum on exiting the EU. It comprises a network with
7,589 users, 532,459 directed follow relationships, and 19,963 tweets, each associated with a binary
stance. The dataset is preprocessed according to [298] to assign each user a scalar value between 0
and 1, referred to as opinion, representing the average stance of the tweets retweeted by the user.
The stance of each tweet is either 0 (“Remain”) or 1 (“Leave”).
- Diginetica[31] [427]: This dataset comprises user session logs from an e-commerce search engine.
The data spans six months and captures user interactions, including clicks, product views, and
purchases. Each user session, defined by a one-hour inactivity period, contains anonymized user
IDs, hashed queries, product descriptions, metadata (price, hashed product names, image identifiers,
and product categories), and log-scaled prices.
- Yoochoose[32] [427]: The Yoochoose dataset contains session data from an online European retailer.
Each session records the user’s click events, with some sessions also including purchase events.
Collected over several months in 2014, the data captures user interactions with the retailer’s website.
The product meta-data includes categories.
**6.6.2** **Tools**
- X-Developer Platform[33]: The X Developer Platform provides powerful tools and resources for
developers to integrate X’s real-time, historical, and global data into their own applications. With
three main products—X API, X Ads API, and X for Websites—the platform supports a wide range
of use cases, from retrieving and analyzing Tweets to managing ad campaigns and embedding X
content directly into websites. The X API offers endpoints for managing conversations, exploring
trends, and engaging with users, while the X Ads API enables businesses to manage ads with
custom targeting and analytics. X for Websites allows seamless embedding of live content to
enhance website engagement. Through comprehensive documentation, libraries, and community
support, the X Developer Platform enables developers to create innovative solutions using X’s data
and engagement tools.
- Reddit-API[34]: Reddit is a news aggregation and discussion platform where posts are organized
into "subreddits," user-created boards moderated by the community. The Reddit API provides
developers access to the site’s extensive collection of posts and comments. This free API has
enabled the development of moderation tools, third-party applications, and training datasets for
LLMs such as ChatGPT, Google, and Gemini. By using this API, we can query tremendous user
interactions (i.e., comments and posts) and their corresponding textual contents (i.e., subreddit
threads), which serve as golden resources for GraphRAG.
- Rec-Bole[35] [553]: RecBole is an open-source, unified, and comprehensive library designed for developing and benchmarking recommendation algorithms. Built with Python and PyTorch, RecBole
offers researchers a streamlined, efficient framework to experiment with over 100 recommendation
algorithms across four main types: General, Sequential, Context-Aware, and Knowledge-Based
Recommendations. The platform simplifies data handling by providing pre-processed copies of
43 benchmark datasets, making it easy for users to dive into model testing and development. The
provided user-product interactions, as well as text-formatted user/product meta-data, could also be
used for GraphRAG.
### 7 Planning and Reasoning Graph
A planning or reasoning graph characterizes the inherent logical flow among different entities, where
entities typically represent concrete planning or reasoning substeps, and edges denote their logical
relations. For the planning graph [356, 569], a common example is a set of API tools used to
achieve certain goals, where nodes represent actions, and edges denote their relational dependencies. For reasoning graphs, a notable example is the recent proposed chain/tree/graph of thoughts
[30https://github.com/somethingx01/TopicalAttentionBrexit?tab=readme-ov-file](https://github.com/somethingx01/TopicalAttentionBrexit?tab=readme-ov-file)
[31https://competitions.codalab.org/competitions/11161#learn_the_details-data2](https://competitions.codalab.org/competitions/11161#learn_the_details-data2)
[32https://www.kaggle.com/datasets/chadgostopp/recsys-challenge-2015](https://www.kaggle.com/datasets/chadgostopp/recsys-challenge-2015)
[33https://developer.x.com/en/docs/x-api/getting-started/about-x-api](https://developer.x.com/en/docs/x-api/getting-started/about-x-api)
[34https://support.reddithelp.com/hc/en-us/articles/16160319875092-Reddit-Data-API-Wiki](https://support.reddithelp.com/hc/en-us/articles/16160319875092-Reddit-Data-API-Wiki)
[35https://recbole.io/](https://recbole.io/)
-----
techniques [24, 437, 488] where each node represents a decision-making thinking step connected by
the reasoning flow. The dependency constraint and reasoning flow in the planning/reasoning graphs
can be naturally represented as relational knowledge, which forms the foundation for GraphRAG in
fulfilling planning/reasoning tasks. This section reviews GraphRAG for the planning and reasoning
graph [24, 248, 282, 342, 345, 356, 355, 368, 372, 454, 437, 477, 488, 555, 569, 142].
**7.1** **Application Tasks**
The representative tasks conducted on planning and reasoning graphs are summarized as follows:
- Sequential Plan Retrieval [356, 355, 372, 454, 555]: As one of the most frequently encountered
tasks, plan retrieval aims to retrieve the plan of steps or tools in the format of subgraphs to complete
user queries. For example, given the user query "Please generate an image where a girl is reading a
book, and her pose is the same as the boy in "example.jpg," then please describe the new image
with your voice.", the retrieved final plan from the global plan graph would be "Post Detection" →
"Pose-to-Image" → "Image-to-Text" → "Text-to-Speech."
- Naturalistic Asynchronous Planning [248]: In contrast to plan retrieval, which considers only
dependency constraints among plans, incorporating time constraints introduces a greater challenge.
Naturalistic Asynchronous Planning aims to produce a plan that meets dependency requirements
and optimizes task completion efficiency, using time summation, time comparison, and constrained
reasoning. For example, a user might request, "To make calzones, here are the steps and times
required; please calculate the optimal plan for completion." An efficient plan would execute "roll
dough," "add filling," and "fill dough" sequentially, with "preheat oven" in parallel, and then
conclude with "bake."
- Structured Commonsense Reasoning [342]: Given a belief and an argument, structured common
sense reasoning aims to infer the stance and generate/retrieve the corresponding commonsense
explanation graph that explains the inferred stance.
- Defeasible Inference [282]: Defeasible Inference is a mode of reasoning in which, given a premise,
a hypothesis may be weakened or overturned in light of new evidence. A prominent approach is to
support defeasible inference through argumentation by constructing inference graphs.
- Tool Usage [570, 142]: Instructing LLMs to use external tools for complex real-world problems
has gained increasing importance. Recent research has explored advanced planning strategies to
enhance LLMs’ tool-use intelligence. Notably, two approaches employ A* search [570] and Monte
Carlo Tree Search [142], both utilizing graph-structured reasoning to adaptively retrieve the next
tool based on the LLM’s internal evaluations and environmental feedback. These methods enable
dynamic tool retrieval, refining the model’s problem-solving precision and flexibility.
- Embodied Planning [368, 477]: Embodied planning tasks in Embodied AI involve guiding agents
to perform sequences of actions based on natural language instructions and visual cues in simulated
or real-world environments. These tasks, such as organizing or cleaning, challenge agents due to
ambiguous instructions, limited task-specific knowledge, sparse feedback, and complex, variable
action spaces.
**7.2** **Reasoning and Planning Graph Construction**
Most existing methods for constructing reasoning and planning graphs begin by analyzing relational
dependencies and subsequently adding edges based on these hard-coded rules. Therefore, rather
than limiting our focus to reviewing this only rule-based construction method, we review various
dependency categories used for edge addition.
- Resource Dependency [355, 372, 454]: This dependency is defined as the shared resources among
different actions/decisions. For example, two tools are connected if the output of one tool matches
the input of the other, enabling a seamless transition from one process to another. The decision to
add edges in existing graph construction methods is made by checking whether one node’s input
matches another node’s output (e.g., plans, tools, or some other abstract processes).
- Temporal Dependency [355]: This relation ensures that the sequence of events follows certain
orders within the planning and reasoning process. For example, connections in some collected
datasets denote the successive order between two APIs for daily life.
-----
- Inclusive Dependency [277]: The dependency described indicates that two connected nodes
belong to the same category or environment. For example, cobblestones and birdhouses are
both part of the category of garden decorations [427]. Hypergraphs can effectively capture such
belonging relationships where one entity belongs to multiple environments [111]. Furthermore,
these dependencies often form hierarchical structures, where "grandparent" entities encompass
"parent" entities, which in turn encompass their "children." As the depth of the hierarchy increases,
the space of possible dependencies grows exponentially, presenting a significant computational
challenge. To address this, many previous works have proposed encoding such hierarchies in
hyperbolic space [277, 257, 527]. To the best of our knowledge, no prior research has explored
inclusive dependencies in RAG systems.
- Causual Dependency [248]: This dependency indicates the cause-and-effect logic within the
graph, where one action/decision causes the trigger of another action/decision. A long-standing
example is the casual graph to encode assumptions about the data-generating process.
- Analogy Dependency [499, 503]: This dependency underscores analogical reasoning, where
relationships take the form "A is to B as C is to D." By recognizing and leveraging such dependency,
humans build on existing knowledge to forge new insights across domains. A powerful historical
example is the discovery of Coulomb’s Law, inspired by the analogy between gravitational forces
affecting celestial bodies and electrical forces between charged particles [322].
While resource-dependent and causal relations involve a sequentially structured relation (former
step/tool/decision leads to the later step/tool/decision), they are inherently different. For instance, if
Tool A generates a report in PDF format and Tool B is designed to extract data from PDF files, Tool
A and B share resource dependency because the output of Tool A (the PDF) matches the input of
Tool B. However, this connection does not imply a direct cause-and-effect relationship between the
two tools, i.e., using Tool A would not necessarily cause us to use Tool B.
**7.3** **Retriever**
Retrievers for handling tasks on reasoning and planning graphs are often modeled as graph traversers.
The query or task instruction locates the initial seeding nodes to initialize the graph traversal; then a
traversal expands the graph’s scope either until a preset budget is exhausted or specific criteria are met.
A core step throughout is selecting the most relevant neighbors from all potential candidates. Based
on the criterion for neighborhood selection, retrievers fall into two main categories: embedding-based
methods, which prioritize neighbors based on embedding similarity, and heuristic-based methods,
which use local and global reward functions to determine neighbor importance.
- Embedding-based: Wu et al. [454] decompose the query, subsequently perform the embedding
similarity match between the concatenated subquery with the current retrieved task API and each
of the existing APIs, and then select the top one from the existing neighboring APIs. It explores the
strategy of training and the one without training.
- Heuristic-based: Compared to embedding-based methods, which rely on dedicated training data
for mapping, heuristic-based methods define rules to guide the graph retriever effectively[555,
569, 142]. Zhuang et al. [569] model tool planning as a tree search algorithm, incorporating A*
search to adaptively retrieve the most promising tool for subsequent use based on accumulated and
anticipated costs. Both cost functions are heuristically designed, drawing on prior literature and
practical insights.
- Thought Propagation Retrieval [499]: Given an input problem, thought propagation retrieval
prompts LLMs to propose a set of analogous problems, and then applies established prompting
techniques, like Chain-of-Thought (CoT), to derive solutions. The aggregation module subsequently
consolidates solutions from these analogous problems, enhancing the problem-solving process for
the original input.
**7.4** **Organizer**
The current GraphRAG literature on planning and reasoning graphs generally omits organizer
mechanisms, as the retrieval process alone achieves sufficient precision, eliminating the need for
reranking. Unlike document or knowledge graphs, which typically apply one-shot embeddingbased similarity retrieval to select the top-K relevant content [452], planning and reasoning graphs
-----
use a multi-round embedding similarity process integrated with reasoning steps, enhancing plan
fidelity. Moreover, reward-based retrieval involves a sophisticated search that further boosts accuracy.
Together, these high-quality strategies reduce the need for fine-grained reranking or filtering.
**7.5** **Generator**
Most existing GraphRAG approaches for reasoning and planning tasks either output the retrieved
plan directly or integrate it into the LLM for downstream solution generation. For instance, Wu
et al. [454] outputs the retrieved graph-structured plan as the final result, while Shen et al. [356]
compiles executed results from expert tools to generate the response. Similarly, after constructing a
tool invocation graph, Shen et al. [355] directly prompts the LLM to generate the parameters, and Lin
et al. [248] leverages the LLM to produce asynchronous plans based on task dependencies, time, and
graph constraints. Notably, most of these works focus on fusing textual information, primarily using
different graph structure formats (e.g., adjacency matrix, adjacency list, edge list, CSR) presented in
textual form.
**7.6** **Resources and Tools**
We summarize the useful resources and tools for GraphRAG on planning and reasoning graphs.
**7.7** **Data Resources**
- Hugging Face[36] [355]: Hugging Face offers a wide array of AI models covering multi-modality
tasks in language, vision, audio, video, and more. Each task corresponds to a tool node that handles
specific input and output. If tools A and B are connected, the output type of A must match the
input type of B. Thus, edges in the Hugging Face plan graph represent the resource-dependent
relation. [355] firstly collects the tool repository and builds a tool graph with a collection of tools
and their dependencies. Then, to generate each question, they sample a subgraph from the tool
graph in the three basic formats: node, chain, and directed acyclic graph (DAG), each of which
embodies a specific pattern for tool invocation. After that, the sampled subgraph is sent to LLMs to
synthesize user instructions and populate the parameters for the tool subgraphs. At last, LLM-based
and rule-based self-critic mechanisms are used to check and filer out the generated instruction to
guarantee quality.
- Multimedia[37] [355]: Unlike the AI-focused tools of Hugging Face, multimedia tools serve a
broader range of user-centric tasks such as file downloading and video editing. The edges remain
consistent with the Hugging Face domain and the tool connections denote the resource-dependent
relation similar to Hugging Face. The construction of Multimedia is similar to Hugging Face.
- Daily Life APIs[38] [355]: Daily life services, including web search and shopping, can also be viewed
as tools for specific tasks. The dependencies among these APIs are primarily temporal, meaning
that two daily life APIs are connected if one follows the other in sequence. The construction of
Daily Life APIs is mostly similar to Hugging Face except that the edges are constructed manually
because of the scarcity of publically available APIs.
- RestBench[39] [372, 454]: A dataset consisting of multiple APIs to address complex real-world user
instructions in two scenarios: TMDB movie database and Spotify music player. The TMDB movie
database offers RESTful APIs encompassing information about movies, TVs, actors, and images.
Spotify Music Player provides API endpoints to retrieve content metadata, receive recommendations, create and manage playlists, and control playback. For RestBench, [372] employed NLP
experts to brainstorm instructions for different combinations of APIs and correspondingly annotate
the gold API solution path for each instruction. Two additional experts are further employed
to thoroughly verify the solvability of each instruction and the correctness of the corresponding
solution path. To adapt RestBench into a graph-structured dataset, [454] treats each API as a unique
task node, modeling their relationships through two key dimensions: (1) categorical association
and (2) resource dependencies. For instance, APIs offering movie-related functionalities, such as
[36https://github.com/microsoft/JARVIS](https://github.com/microsoft/JARVIS)
[37https://github.com/microsoft/JARVIS](https://github.com/microsoft/JARVIS)
[38https://github.com/microsoft/JARVIS](https://github.com/microsoft/JARVIS)
[39https://restgpt.github.io/](https://restgpt.github.io/)
-----
retrieving movie details or recommending films, are grouped under the ’movie’ category, while
APIs focused on person-related tasks, like searching for actors, are classified under the ’person’
category. Additionally, if two APIs share a common parameter (e.g., movie-id), a link is established
to represent resource dependency. To further enhance semantic distinction, GPT-4 is prompted to
assign a unique name and detailed functional description to each API.
- AsyncHow[40] [248]: A curated dataset consisting of 1.6K data points for asynchronous planning.
Each data point comprises a user instruction specifying tasks with their basic execution constraints,
represented by a directed acyclic graph (DAG). In these DAGs, nodes denote actions, and edges
represent ordering constraints. Each edge carries a weight indicating the time required to complete
the preceding action and transition to the next. Additionally, edges also signify causal links, meaning
an action can only proceed if all preceding linked actions are completed. To construct this dataset,
we first collected planning tasks from WikiHow [214]. LLMs were then used for preprocessing,
time annotation, and step dependency annotation. Specifically, plans containing optional steps,
irrelevant tasks, or those lacking quantifiable duration were filtered out. The GPT-3.5 is used to
estimate the time duration for each step, and the GPT -4 is used to annotate step dependencies using
the DOT language. To generate natural language questions with execution constraints, ten trivially
different but plausible templates were used to paraphrase the graph-structured dot language into
human-understandable texts. The optimal time duration for a plan was calculated by determining
the longest path within the DAG representation of the workflow.
- EXPLAGRAPHS[41] [342]: Give the initially collected triplets consisting of belief, argument, and
stance, the commonsense explanation graphs are constructed through a generic create-verify-refine
iterative framework. Firstly, annotators construct a commonsense-augmented explanation graph
that explicitly explains the stance. Each graph comprises 3-8 facts, each of which is a triplet
with two concepts as entities connected by their relations. The graphical representation allows us
to automatically perform in-browser checks for structural constraints, thereby guaranteeing the
structural correctness of the graph. To further ensure the semantic correctness of the constructed
graph, annotators reason through the graph to infer the stance based solely on the belief and the
explanation graph. Finally, for incorrect graphs, another annotator refines them either by adding a
new fact, removing an existing one, or replacing an existing fact.
- GSM8K[42] [65]: The GSM8K dataset is a collection of 8.5K high-quality, linguistically diverse
math word problems designed to assess multi-step reasoning for question answering. These grade
school-level problems, solvable by a bright middle school student, require between 2 and 8 steps,
primarily using basic arithmetic operations (+, −, ×, ÷) without needing concepts beyond early
Algebra. Solutions are presented in natural language to facilitate general applicability, offering
insight into the reasoning processes of large language models. This format highlights how models
handle structured, step-by-step reasoning in response to real-world math problems.
- PrOntoQA[43] [348]: PrOntoQA is a question-answering dataset that provides examples featuring
chains-of-thought to outline the reasoning needed for correct answers. The sentences are syntactically simple and well-suited for semantic parsing, making it valuable for formal analysis of
predicted reasoning chains from large language models like GPT-3.
**7.7.1** **Tools**
- ToolBench[44] [355]: Recent studies on software tool manipulation with LLMs primarily depend
on closed model APIs (e.g., OpenAI), as there remains a significant performance gap between
these proprietary models and available open-source LLMs. To investigate the underlying causes
of this discrepancy and to advance the capabilities of open-source LLMs—particularly in tool
manipulation—a benchmark named ToolBench has been developed. ToolBench includes a range of
diverse software tools designed for real-world tasks and provides an accessible infrastructure for
directly evaluating each model’s execution success rate.
[40https://github.com/fangru-lin/graph-llm-asynchow-plan](https://github.com/fangru-lin/graph-llm-asynchow-plan)
[41https://github.com/swarnaHub/ExplaGraphs](https://github.com/swarnaHub/ExplaGraphs)
[42https://github.com/openai/grade-school-math](https://github.com/openai/grade-school-math)
[43https://github.com/asaparov/prontoqa](https://github.com/asaparov/prontoqa)
[44https://github.com/sambanova/toolbench](https://github.com/sambanova/toolbench)
-----
### 8 Tabular Graph
Tabular data is another type of structured data that is widely used in real-world applications [343] and
is typically stored in relational databases [66]. Tabular data may consist of a single table containing
samples and their attributes, or multiple tables that share primary and foreign keys. LLMs have been
explored to process and solve tasks involving tabular data, primarily by transforming the tabular data
into text through serialization [107, 365]. However, such serialization may lead to some issues: (1) In
the case of a single table, the feature columns should exhibit permutation invariance, but serializing
the table data may disrupt this invariance; (2) For multiple tables, one table may be connected to
another through primary/foreign keys, and leveraging these relationships is crucial for various tasks.
Therefore, graph structures can be a suitable representation for tabular data. Additionally, when
tables contain too many rows to fit within the context window of LLMs, graphs can facilitate efficient
retrieval.
**8.1** **Application Tasks**
Tables are widely used in real-world scenarios to store features and relationships between data.
Understanding the structure of tabular data is crucial. As a result, there are numerous tasks that can
benefit from the use of tabular graphs and below we list a few representative ones.
- Node-level tasks: Node-level tasks include node classification and node regression, applied to tasks
like cell type prediction [188], fraud detection [335, 364], outlier detection [124], and click-through
rate (CTR) prediction [242, 94].
- Link-level tasks: Link-level tasks involve link prediction, edge classification, and edge regression.
Many tabular data tasks can be modeled as link-level tasks, such as data imputation [557, 497] and
recommendation [339].
- Graph-level tasks: Graph-level tasks aim to predict the properties of the entire graph, such as table
type classification and table similarity prediction [432, 178, 188].
- Table question answering: Table QA involves generating answers by understanding and reasoning
over tabular data. This task requires comprehension of both the content and their relationships
within tables, making graph structures suitable for encoding such information. For example, Zhang
[532], Zhang et al. [531] utilize graphs to enhance Table QA.
- Table retrieval: Table retrieval focuses on retrieving semantically relevant tables based on natural
language queries [411, 61].
**8.2** **Tabular Graph Construction**
Graphs are used in tabular data learning to model high-order feature interactions, high-order instance
relationships, and relationships between instances across multiple tables. There are typically two
types of nodes: instance nodes, which represent each row of a table, and feature nodes, which
represent individual features. Generally, the following graphs are constructed:
- Instance Graph: An instance graph connects a table’s rows (instances), modeling the relationships
between instances. It is particularly useful for retrieving relevant instances within tabular data.
There are mainly two approaches for constructing instance graphs:
**– Rule-based methods: Instances are connected based on predefined rules. For example, heuristics**
derived from expert knowledge can be used to connect instances that share certain features [267].
Expert knowledge can be leveraged for the graph construction [314]. Another common approach is similarity-based methods, such as connecting instances through K-Nearest Neighbors
(KNN) [128, 102] or based on similarity exceeding a threshold [374, 44].
**– Learnable methods: In this approach, an instance graph is typically initialized using rules or**
heuristics. The edge weights are then adjusted during the learning process, allowing the model
to dynamically refine the graph structure over time [194, 245, 67, 197].
- Feature Graph: A feature graph connects features, with edges representing correlations between
pairs of features. Typically, feature relationships are modeled in a learnable manner [480, 559].
Some works construct the feature graph by leveraging feature similarity [136, 242], while others
-----
use heuristics based on expert knowledge to establish connections [223]. Additionally, certain
approaches link features if they belong to the same instance [335].
- Instance-Feature graphs: The instance-feature graph is a heterogeneous graph, which connects
the instance with their corresponding features [135, 497, 411, 451].
- Cell Graph: Cell graphs treat each cell in the table as a node. Xue et al. [479] build a cell graph,
where each cell node contains both spatial and logical locations attributes while the adjacency
matrix represents the neighbor relation or the same-row and same-column relation between two
cells.
- Tabular Hypergraph: A hypergraph is a generalization of a graph in which an edge, known as a
hyperedge, can connect multiple nodes. Since tables are often invariant to arbitrary row and column
permutations, a hypergraph is a suitable structure for modeling tabular data [506]. Specifically,
Chen et al. [46] model each row and column in a table as a hyperedge for pretraining purposes.
Additionally, Du et al. [94] construct a hypergraph from tabular data to capture relationships
between instances effectively.
Additionally, Cvetkov-Iliev et al. [70] model table as a knowledge graph, where the feature values
or rows represent nodes, and the column names represent the relations. Zhang et al. [531] build a
heterogeneous graph based on various predefined relations. Wang et al. [432] construct trees for
non-relational tables based on the hierarchical relations.
Previous approaches primarily focus on constructing graphs based on a single table. However, in
relational databases, there are often multiple tables. A common method is to first merge these tables
into a single table, and then apply the previously mentioned methods to the merged table [195].
However, this approach relies on manually joining the tables as part of a feature engineering
step, which requires significant effort and substantial domain expertise. In the following, we will
introduce cross-table graphs, which directly build a graph from multiple tables without the need for
manual merging.
- Cross-table Graphs: Cross-table graphs for tabular data typically connect different instances
across multiple tables, where nodes usually represent rows in each table and edges are defined
by primary-foreign key relationships [115, 71, 517]. Additionally,Wang et al. [417] propose
Row2N/E: if a table has a primary key (PK), each row is treated as a node; however, if there are
also two foreign key (FK) columns, each row is instead represented as an edge. These graphs
model relationships across multiple tables, enabling the integration of information from different
perspectives to facilitate various tasks.
Additionally, Bai et al. [15] build a hypergraph based on multiple tables, where the primary or
foreign key as nodes, and the nodes within the same table as hyperedge.
**8.3** **Retriever**
Modeling tabular data into graphs is still in its early stage, and most works follow popular retrieval
methods, as introduced in Sections 2.4. Additionally, Fey et al. [115] retrieve subgraphs based
on timestamps to construct time-consistent computational graphs, while Cvitkovic [71] adopt a
deterministic heuristic to select subgraphs.
**8.4** **Generator**
Most existing methods that leverage tabular graphs still rely on GNNs or Graph Transformers
as generators [223]. Besides, Wang et al. [417] fuse both GNNs and tabular based predictors,
such as DeepFM [129], FT-Transformer [125], XGBoost [50] and AutoGluon [101]. Ivanov and
Prokhorenkova [171] jointly train gradient boosted decision tree (GBDT) and GNN. Recent advancements have seen the application of LLMs to tabular data tasks. This is typically achieved by
converting tabular data into sequential formats suitable for LLM input, as explored by Fang et al.
[107], Dong and Wang [89], and Sui et al. [378]. However, this transformation can lead to the loss
of inherent structural information present in the original tabular format. With advancements in the
application of LLMs to graph data, LLMs can be further enhanced to handle various tasks more
effectively with the support of tabular graphs.
**8.5** **Resources and Tools**
In this section, we list some data sources and tools for GraphRAG on tabular graphs.
-----
**8.5.1** **Data Resources**
Relational machine learning has recently gained popularity, leading to the development of several
benchmarks and datasets for tabular data. For example:
- RelBench [339, 115][45] is a collection of realistic, large-scale, and diverse benchmark datasets for
machine learning on relational databases. It includes various tasks, such as Node-level prediction
tasks (e.g., multi-class classification, multi-label classification, regression), Link prediction tasks,
Temporal and static prediction tasks. The benchmark features several datasets, including relamazon, rel-stack, rel-trial, rel-f1, rel-hm, rel-event, and rel-avito, providing a robust platform for
evaluating models across multiple relational data scenarios and tasks.
- TabGraphs [23] evaluates various models — including simple baselines, tabular models, and GNNs
— on graphs with tabular node features. This benchmark includes several TabGraphs datasets:
tolokers-tab, questions-tab, city-reviews, browser-games, hm-categories, hm-prices, city-roads-M,
city-roads-L, avazu-devices, web-fraud and web-traffic.
- Tabular-benchmark [126] [46] benchmarks 45 tabular datasets from diverse domains, comparing the
performance of several tree-based and deep learning models. They evaluate tasks such as numerical
classification, numerical regression, categorical classification, and categorical regression.
- Shwartz-Ziv and Armon [363] benchmark 11 tabular datasets, such as MSLR [325], Forest Cover
Type, Higgs Boson, and Year Prediction [96]. evaluating several tree-based models, deep learning
models, and ensemble models.
- DBInfer Benchmark [417][47] is a set of benchmarks for measuring machine learning solutions
over data stored as multiple tables. It includes several large-scale relational database (RDB)
datasets—such as AVS, OB, DN, RR, AB, SE, MAG, and SE—with tasks like retention, CTR,
purchase, CVR, churn, rating, popularity, venue prediction, citation, charge, and prepay. The
benchmark provides implementations of various baselines, including popular tabular models with
and without auto-feature-engineering methods, as well as Graph Neural Networks, making it a
robust resource for multi-table learning evaluations.
- RTDL[48]: RTDL (Research on Tabular Deep Learning) is a collection of papers and packages on
deep learning for tabular data. It provides several popular tabular deep learning models.
- OpenML [49] is an open platform for sharing datasets, algorithms, and experiments. It provides
a large collection of machine learning datasets across various domains, including many tabular
datasets.
- Kaggle [50] hosts a wide array of datasets for data science competitions, making it a valuable resource
for benchmarking tabular machine learning models on diverse real-world data.
- UC Irvine Machine Learning Repository [96][51] is a collection of databases, domain theories,
and data generators that are used by the machine learning community for the empirical analysis of
machine learning algorithms.
- HiTab [61] is a hierarchical table dataset for question answering and natural language generation.
**8.5.2** **Tools**
In this subsection, we list several tools that provide essential functionalities for data preparation,
model training, and evaluation. These tools can be effectively leveraged for GraphRAG applications
on tabular graphs.
- PyTorch Tabular [190][52]: PyTorch Tabular is a powerful library built on PyTorch designed to
simplify the application of deep learning techniques to tabular data. It provides several data
45https://relbench.stanford.edu/
46https://github.com/LeoGrin/tabular-benchmark
47https://github.com/awslabs/multi-table-benchmark
48https://github.com/yandex-research/rtdl
49https://www.openml.org/
50https://www.kaggle.com/
51https://archive.ics.uci.edu/
52https://github.com/manujosephv/pytorch_tabular
-----
preprocessing functions, including normalization, standardization, encoding of categorical features,
and dataloader preparation. Additionally, it includes a variety of tabular machine-learning models
and evaluation functions, making it a comprehensive tool for handling tabular data.
- DeepTables [179][53]: DeepTables is an easy-to-use toolkit that harnesses the power of deep learning
for tabular data. Built on TensorFlow, it is designed to address classification and regression tasks
on tabular data. DeepTables offers a variety of tabular models and supports AutoML.
- PyTorch Frame [156][54]: PyTorch Frame is a deep learning extension for PyTorch, designed for
heterogeneous tabular data with different column types, including numerical, categorical, time, text,
and images. t provides a wide range of tabular models and supports integration with diverse model
architectures, including Large Language Models.
### 9 Other Domains
While substantial GraphRAG research has been investigated in previous domains, GraphRAG remains
significantly underexplored in other domains such as infrastructure, biology, and scene. Therefore,
we combine them into a single comprehensive section, focusing only on key studies in these fields.
**9.1** **Infrastructure Graph**
Infrastructure graphs, defined by Points of Presence (PoPs) interconnected through physical links,
play a vital role in serving our daily activities across sectors such as power, water, gas, transportation,
and communication. Due to the scarcity of GraphRAG research in infrastructure graphs, our review
mainly focuses on graph construction methods and tasks alongside only a brief overview of recent
(Graph)RAGs in this domain.
- Power Networks [571, 311, 247]: In power networks, nodes represent critical entities such as
power plants, substations, transformers, and load points, while edges correspond to transmission
and distribution lines facilitating the flow of electrical power across the network. Each node
possesses features like power generation capacity, voltage level, load demand, geographical location,
reliability metrics, and operational status, capturing the network’s operational and spatial aspects.
Edges are characterized by features such as transmission capacity, impedance, length, voltage level,
and real-time power flow, all essential for analyzing line performance and flow efficiency. This
graph-based representation enables a range of essential tasks, including transformer fault diagnosis,
power outage prediction, power flow approximation, and power generation optimization.
- Water Networks [465]: In water networks, nodes are junctions (pipe connections or users) and
sources (reservoirs and tanks), and edges are pipes with water flow. Basic tasks include estimating
node water heads and flow in all pipes given the water network layout, pipe characteristics, nodal
demands, reservoir levels, and head measurements at limited locations [32].
- Gas Networks [407, 486]: nodes represent connection points, gas sources, storage facilities,
compressor stations and users, and edges edge respected natural gas pipelines, carrying user loads,
diversion, and input.
- Transportation Networks [424, 184]: Road networks, where intersections serve as nodes and road
connections as edges, are commonly used to model spatial dependencies for traffic flow and speed
forecasting. Similarly, metro and bus networks use stations as nodes with routes as edges, capturing
station topology. Region-level problems involve dividing cities into regular or irregular regions,
represented as graph nodes, where spatial dependencies reflect land use patterns; regular regions
often use grid partitions, while irregular ones use diverse methods like road- or zip-code-based
partitions. At the road level, sensor, segment, intersection, and lane graphs capture different road
elements, with nodes representing sensors, segments, intersections, or lanes. For station-level
networks, nodes represent various transportation hubs (subway, bus, bike, or car-sharing stations)
linked by natural connections like subway or road networks, creating an interconnected spatial
dependency graph.
53https://github.com/DataCanvasIO/DeepTables
54https://github.com/pyg-team/pytorch-frame
-----
- Communication Networks [113, 341]: A computer network system typically comprises two
interconnected graph layers: the logical and the physical. The logical layer represents the flow of
data, where traffic is routed through multiple intermediate routers before reaching its destination.
In this layer, nodes correspond to routers, and edges represent the logical paths that data packets
traverse. Conversely, the physical topology layer refers to the underlying infrastructure, with Points
of Presence (PoPs) from providers such as Comcast and Verizon serving as nodes and the physical
fiber links between them forming the edges. Studies focused on the logical layer often model
networks as graphs, where nodes represent devices (e.g., switches, routers) and edges denote links
(e.g., fiber-optic cables) or traffic paths [177]. This graph-based representation captures both the
forwarding behavior of the network (i.e., traffic interactions) and its connectivity (i.e., topological
behavior), offering valuable insights for network management. For instance, optical topology-aware
traffic engineering has proven effective in mitigating issues like fiber cuts and flash crowds. For the
physical layer, much previous research has focused on inferring the complete as well as aligning
the physical and logical layers.
To summarize, the tasks operated on infrastructure networks could be broadly encompassed as
utility prediction (e.g., flow and node performance), flow simulation and generation, vulnerability
analysis, and network maintenance and operation. Managing and understanding the complex physical
relationships within these graphs is essential for improving infrastructure management in key areas
like service delivery, function forecasting, and network optimization. In computer networks, for
instance, the state of a logical traffic routing path is directly influenced by the underlying physical
fiber links. Predicting critical metrics like traffic delay and jitter depends on assessing the status
of all involved fiber links [113, 341], as these metrics are highly interdependent. The intricate
physical connections in infrastructure graphs create valuable opportunities for employing GraphRAG
techniques to enhance infrastructure management and optimization. Although no existing works have
explored GraphRAG in infrastructure, few works have leveraged RAG, which we briefly review in
the next.
Hussien et al. [168] explore the use of RAG to empower automated vehicles with the ability to
anticipate pedestrian and driver behaviors, including pedestrian road-crossing actions and driver lane
changes. RAG retrieves explanatory documents from the JAAD and PSI datasets to provide insights
into pedestrian behavior. Additionally, Qian et al. [323], Wang et al. [418], Liu et al. [250], Bariah
et al. [20] build on recent advances in generative models for vision and language, proposing to
harness the capabilities of LLMs within computer networking. A systematic approach has been
proposed to develop a foundation model for traffic tasks, such as traffic classification and generation,
treating packet hexadecimal bytes as tokens. Furthermore, the potential of generative AI in advancing
telecom and networking is reviewed. Specifically, Kotaru [213] introduces an "operator’s copilot,"
a natural language interface that leverages LLMs for efficient data retrieval. This interface helps
manage thousands of counters and metrics, minimizing the need for specialists and accelerating
issue resolution through data intelligence. These pioneering works in leveraging generative AI for
computer networking lay a good foundation for rebuilding (Graph)RAG for infrastructure networks.
**9.2** **Single-cell Graph**
Single-cell sequencing allows for the detailed analysis of molecular traits at the level of individual
cells. For example, single-cell RNA sequencing (scRNA-seq) [210] quantifies RNA transcript levels,
offering valuable information about cell identity, developmental stages, and functional characteristics.
Single-cell assay for transposase-accessible chromatin sequencing (scATAC-seq) [29] records the
number of reads per accessible chromatin region, resulting in highly dimensional data matrices
containing hundreds of thousands of genomic regions. With the explosion of the number of singlecell data, a variety of deep learning approaches [299, 85, 387, 426, 442] have been developed in
recent years to tackle single-cell analysis, especially GNN methods [268, 84, 441, 352, 422] have
been applied to various downstream tasks. In the task of cell type annotation, sigGCN [422] builds
a gene-wise weighted adjacency matrix using data from the STRING database [385] to establish a
gene interaction network, with the node features representing corresponding gene expression levels.
This graph is fed into a GCN-based autoencoder, which includes a convolutional layer and a maxpooling layer, followed by a flattened layer and a fully connected (FC) layer, to perform cell type
annotation. scDeepSort [352] constructs a weighted bipartite graph where both cells and genes serve
as nodes, with the edge weights representing the gene expression values for each cell-gene pair. Gene
-----
node features are derived through principal component analysis (PCA), while cell node features are
obtained by aggregating the weighted features of the connected gene nodes. The most common way
to construct a single-cell graph is to build a KNN graph from single-cell data [414]. To be specific, the
data is first normalized to make it comparable across different cells. Next, dimensionality reduction
methods, such as principal component analysis (PCA) [281], are employed to reduce the dataset’s
high dimensionality while retaining key information. In this lower-dimensional space, distances
between cells are calculated, often using Euclidean distance. Each cell’s K nearest neighbors are then
determined, and a graph is formed where cells are represented as nodes, and edges connect each cell
to its K closest neighbors.
**Multi-omics single-cell graphs: Multi-omics single-cell technologies [22] integrate multiple layers**
of biological information—such as gene expression (RNA), chromatin accessibility (ATAC), and
protein data—at the resolution of individual cells. Multiome [375] refers to the simultaneous
measurement of multiple molecular modalities, such as gene expression (RNA) and chromatin
accessibility (ATAC), from the same single cell while CITE-seq (Cellular Indexing of Transcriptomes
and Epitopes by Sequencing) [295] is a multi-omics technology that allows for the simultaneous
measurement of gene expression and protein surface markers at the single-cell level. Taking multiome
dataset as an example, scMoGNN [441] builds a heterogeneous graph consisting of cell nodes, gene
nodes, and peak nodes. The edge between a cell node and a gene node indicates the expression level
of that gene in the cell, while the edge between a cell node and a peak node reflects the expression
level of that peak in the cell. Additionally, a threshold can be applied to filter and select the connecting
edges. Additionally, pathways and gene activity can be incorporated as prior knowledge to establish
connections between gene nodes, and connections between peaks and gene nodes respectively. This
graph representation captures both intra-modality (within each modality, like gene expression or
chromatin accessibility) and cross-modality (between gene expression and chromatin accessibility)
relationships, providing a comprehensive view of regulatory dynamics at the single-cell level
**Spatial transcriptomics single-cell graphs: Unlike traditional transcriptomics, spatial transcrip-**
tomics retains the precise location of spot representing a cell or small group of cells within the
tissue. Spatial transcriptomics primarily consists of three key data sources: gene expression, which
captures transcript levels in specific spots; spot location, providing the spatial coordinates of each
spot within the tissue; and spot images, offering visual context and morphology of the tissue at each
spot. In the task of cell type deconvolution, when building the KNN graph, GNNDeconvolver [84]
integrates both cell position information and gene expression data to compute the similarity between
cells. SpaGCN [154] additionally incorporate spot morphology to calculate similarity between spots.
Together, constructing KNN graph with such three data sources provides a richer, more holistic view
of the tissue, enabling deeper insights into cellular function, interactions, and spatial organization.
**9.3** **Scene Graph**
Scene graphs are a data structure designed to capture spatial and semantic relationships between
objects within a scene. They are widely used in computer vision, graphics, and robotics to organize
scenes, particularly when multiple objects interact, or the scene contains complex interactions. For
example, SceneGraphs [147] is a dataset for visual question answering, containing 100,000 scene
graphs that describe the objects, attributes, and relationships within images. It presents tasks that test
spatial reasoning and multi-step inference by asking users to answer open-ended questions based on
textual descriptions derived from scene graphs.
G-Retriever [147] processes Scene Graphs by first parsing JSON data to index objects and attributes,
allowing for efficient retrieval of relevant nodes and edges based on the query. It then constructs a subgraph with the necessary scene information, filtering out unrelated details. Finally, G-Retriever uses
this subgraph along with an LLM to generate answers, leveraging spatial and semantic relationships
for accurate reasoning and response generation.
P-RAG [477] engages with the environment through agents, building and updating a database of
historical trajectories that guide the agent’s actions. Initially, the task’s goal instruction is provided.
Before each action, the agent captures an observation image reflecting the current state, which is then
converted into a scene graph format to facilitate processing by LLMs. By creating and retrieving
scene graph context, P-RAG enhances the LLM’s ability to accurately interpret and navigate complex
scenes, making this approach especially useful in planning embodied everyday tasks.
-----
**9.4** **Random Graph**
Random graphs are a foundational concept in network theory [304, 174] and are widely used for
modeling and studying complex networks in fields such as computer science, physics, biology,
and social sciences. They are constructed based on probabilistic rules, leading to various possible
structures that reflect the diversity seen in real-world networks. These random graphs can be used to
analyze the retrieval, organization, and generation processes within GraphRAG.
There are many models to generate random graphs, such as Erd˝os-Rényi Model (ER Model) [100],
Watts–Strogatz model [436], Barabási–Albert model [7], Random geometric graph [33], scale-free
networks [18] and stochastic block model [150]. Additionally, various graph structures, such as
Path Graphs, Complete Graphs, Star Graphs, and Barbell Graphs, can be generated to meet specific
analytical needs.
Random graph generation models, such as Contextual stochastic block models (CSBM) [80] has been
widely leveraged in GNNs analysis [19, 280, 285, 140]. Additionally, random graphs are increasingly
applied to examine the behavior of LLMs. For instance, Fatemi et al. [109] leverage Erd˝os-Rényi
graphs, scale-free networks, Barabási–Albert, stochastic block model, star, path and complete graph
generators to test LLM performance on various graph reasoning tasks, such as edge existence, node
degree, node/edge count, connected nodes and cycle check. GraphLLM [36] generates random
graphs in various input formats to assess LLM performance on graph reasoning tasks, including
substructure counting, maximum triplet sum, shortest path calculation, and bipartite graph matching.
In this approach, a graph transformer encodes the graph, and the embedding-fusion method is used
for generation. Bachmann and Nagarajan [13] use star graphs to investigate limitations within the
next-token prediction paradigm. Other studies, such as Dai et al. [73], Wang et al. [413], Guo et al.
[130], and Luo et al. [275], explore the graph reasoning abilities of LLMs across various tasks.
### 10 Challenges and Future Work
After proposing the holistic framework of GraphRAG and reviewing it in each domain, we begin in
this section by outlining the challenges and opportunities associated with each key component of
GraphRAG, including graph construction, retriever, organizer, and generator. Then, we discuss the
challenges and opportunities for GraphRAG as a holistic system with its evaluation and application.
**10.1** **Graph Construction**
- How to construct graphs? There are numerous ways to construct graphs, yet different tasks
or domains may require different graph structures. For example, deciding on the granularity of
nodes and edges, as well as which entities or relationships to extract, is critical. This process may
also need to address challenges such as entity disambiguation, entity alignment, and coreference
resolution. Understanding when a graph structure is necessary, whether to use single or multiple
graphs and how to construct them in the most appropriate way for a specific application is essential
but often complex.
- The format of graphs. Graphs can be represented in various forms—are these representations
equivalent, or do they offer unique advantages? Choosing the most effective representation for a
given task can significantly influence performance.
- Multi-modal Graphs. Despite building text-based graphs, the retrieved resources, such as images,
audio, or video, can be multi-modal. Constructing a cohesive graph from multi-modal data presents
a significant challenge, as it requires integrating diverse data types while preserving meaningful
relationships.
- Dynamic graphs. Many real-world scenarios involve dynamic data that evolve over time, which
are essential for downstream tasks. Developing strategies to construct, update, and store graphs
dynamically while maintaining both efficiency and effectiveness presents a further challenge worthy
of exploration.
-----
**10.2** **Retriever**
- Differentiating Neural and Symbolic Knowledge: Graph-structured data usually involves two
distinct types of knowledge: symbolic-formatted knowledge, such as relations in knowledge graphs,
and neural-formatted knowledge, such as entity names. Developing techniques to differentiate
these two types of knowledge and designing corresponding retriever strategies for retrieving these
two types of knowledge is worthy of further investigation.
- Harmonization between internal and external knowledge: Since GraphRAG is often applied
when internal knowledge alone is insufficient to address the task, assessing any overlap between
internal and external knowledge using a robust calibration method is crucial. In addition, external
knowledge may sometimes conflict with internal knowledge. To handle this, it is essential to design
an effective knowledge validation and reconciliation mechanism, allowing selective retrieval and
updating of external content as needed.
- Trade-off among Accuracy, Diversity, and Novelty: Real-world user retrieval often involves
complex intentions. Beyond delivering accurate content to ensure high utility—such as achieving
high question-answering accuracy—there may also be a demand for diversity and novelty in the
retrieved information. Balancing retrieved content’s accuracy, diversity, and novelty remains an
open-ended challenge.
- Reasoning, planning, and thinking along the way: A real-world retriever may need to dynamically and adaptively update its retrieve process in response to both the initial query and the content
retrieved along the way. The question of how to equip the retriever with these adaptive thinking,
reasoning, and planning abilities remains a big problem.
**10.3** **Organizer**
- Balancing Completeness and Conciseness. The retrieved graphs may be large, potentially
containing significant information that is not related to the query. The Organizer should balance the
need for complete information with the risk of overwhelming the model. This requires techniques
for pruning irrelevant nodes and edges while retaining essential context, which is especially
challenging in large or noisy graphs. Additionally, some knowledge may already be captured
by LLMs, raising the question of how to identify and remove redundant information to further
improve efficiency.
- Optimal Data Structuring: Determining the most effective way to structure the retrieved content
is a challenge. For example, deciding how to convert a structured graph into a format that the
generator can leverage, how to arrange the order of retrieved content, and how to preserve the
original data structure for structure-sensitive tasks are all important considerations. Different tasks
may benefit from different structuring methods.
- Aligning Different Resources: Retrieved content may come from various sources and in diverse
formats, such as text, graphs, and images. Aligning these components effectively to help the
generator poses a significant challenge, especially when integrating multiple data modalities.
- Data Augmentation: Incorporating data augmentation within the Organizer involves enriching
the retrieved graph content with nodes, edges, or features to improve the robustness of the model
and improve downstream tasks. However, this process requires balancing real and augmented data
to avoid introducing irrelevant or redundant information.
**10.4** **Generator**
- Correct Format for Prompting. The content retrieved after organization varies significantly in
format — such as texts, triplets, or graphs — while current LLMs can only process text inputs.
Exploring the most effective format for optimal LLM performance on specific tasks and designing
more flexible generators that go beyond text input could be worthwhile avenues for research.
- Structural Encoding. When the retrieved content is a subgraph and the downstream generator
is an LLM, ensuring that the LLM can interpret the graph’s structural information is essential.
Designing effective structural encodings and integrating them into token embeddings pose a key
challenge. Although some previous works have incorporated various graph encodings into the
-----
text decoding process, no systematic study has yet demonstrated whether LLMs can distinguish
between these encodings and accurately recognize their corresponding geometric structures.
**10.5** **GraphRAG as a System**
Rather than individual components, GraphRAG is a system. Designing an efficient and cohesive
GraphRAG system presents additional challenges:
- Integration Across Components: Ensuring seamless interaction among the Graph Construction,
Retriever, Organizer, and Generator components is essential. Each part must operate harmoniously
to maintain efficiency and accuracy, which can be challenging to optimize the system globally.
- Scalability: As data volumes and query demands increase, each component of the GraphRAG
system, such as Graph Construction, Retriever, Organizer, and Generator, must efficiently handle
larger, more complex graphs without compromising performance.For example, efficient graph
storage, optimized querying (e.g., subgraph sampling and pathfinding), streamlined organization
of retrieved components, and responsive generation are all essential. Additionally, training,
serving, fine-tuning, and evaluating within GraphRAG demand sophisticated engineering on
modern hardware and software stacks, including efficient training, data-efficient fine-tuning,
communication-efficient algorithms, implementation of reinforcement learning with human
feedback, GPU acceleration and other specialized hardware, model compression for deployment,
and online maintenance.
- Trustworthiness: Real-world GraphRAG systems often operate in high-stakes domains such
as education [394], healthcare [466, 507], and law [446], which imposes multiple desires when
deploying GrapRAG systems. Therefore, deeply understanding the broader desires beyond merely
optimizing utility is essential to ensure effective deployment. As motivated by trustworthy and
safety research in other fields [39, 166, 420], these key goals beyond utility include reliability,
robustness, fairness, and privacy. Although previous works have initiated the exploration of this
multi-purpose optimization of RAG [405, 561], they mostly focus on RAG without dedicated
investigation of how the relational information captured by GraphRAG could cause additional
trustworthy and safety concerns.
**– Reliability [215, 328, 376, 495, 235, 305]: Reliability requires the system to deliver consis-**
tently low error rates across different scenarios. In RAGs, uncertainty quantification presents
two primary challenges: Firstly, during the generation phase, uncertainty stems from the
inherent probabilistic generation of LLMs. Standard techniques such as conformal prediction
have been applied here with few adaptations [215, 328, 376, 495]. The general idea is to
estimate the non-conformal score based on the calibration set and filter out those high-risk
answers during the testing phase. Secondly, the retriever and its interaction with LLMs
(e.g., treating LLMs as the reasoning agent to perform adaptive retrieval) also possess uncertainty [235, 305]. Unlike RAGs, retrievers and generators in GraphRAG demonstrate the
multi-hop nature, which raises new requirements for multi-stage uncertainty quantification.
For example, the error rate of one-step generation might accumulate after multi-step generation, and how to calibrate this accumulated error at the global level is the key challenge. The
very first work [305] proposed the learn-then-test framework to guarantee the global error
rate. However, this solution may not consider the uncertainty of human-LLM interactions.
Moreover, the additional learn-then-test component takes additional time and computational
load for calibration. Future research aims to develop methods that can quantify and calibrate
the uncertainty of (Graph)RAG as a system holistically.
**– Robustness [105, 460, 474, 496, 97, 542, 564]: Robustness aims the system to deliver equal-**
quality response under extreme scenarios such as the presence of noisy and irrelevant content.
Existing research on RAGs has highlighted the significant degradation in LLMs’ performance
when retrieved content includes noise or irrelevant information [496, 105]. Some studies
address this challenge by adversarially training LLMs in noisy environments. However, there
has been limited exploration of LLMs’ robustness from the structural perspective. For example,
would the LLM reasoning performance change when the underlying reasoning graph gets
perturbed [97, 542, 564]? Addressing these issues requires a deeper focus on the interplay
-----
between structural integrity and the robustness of GraphRAG systems, paving the way for
robust performance in tasks requiring complex reasoning and planning.
**– Safety [573, 478, 435, 79, 460]: Recent studies have highlighted that LLMs are susceptible**
to various adversarial attacks. Techniques such as prompt manipulation, hint injection, and
input perturbation enable attackers to bypass safety mechanisms and exploit vulnerabilities, posing substantial social risks. RAGs that combine the power of LLMs with external
databases introduce unique safety challenges. Existing works in RAGs have developed several
threats (e.g., adversarial passage injection [573], group-query targeting [478], and jailbreak
attacks [435, 79]) and defense strategies (such as isolate-then-aggregate strategy [460]). However, all of them have overlooked the structural vulnerabilities inherent in graph-structured
data. For instance, attackers can exploit the network structure of graph-based databases by
leveraging principles from network science to design highly impactful attacks. Attackers
can maximize their influence on the system’s behavior by strategically targeting nodes with
specific structural properties, such as high degree or centrality. This highlights the need for
robust defenses that account for both content-based and structural threats in RAG systems.
**– Privacy [41, 505, 324]: Privacy is a critical concern in RAG systems [513, 561], especially**
when operating in domains involving sensitive or personal data, such as healthcare, education,
or finance [489, 74, 301]. Unlike traditional RAG systems, GraphRAG introduces additional
privacy risks due to the relational nature of graphs, which may inadvertently reveal private
information through connections or patterns. For instance, even if a sensitive node is protected,
it may still be retrieved indirectly via its neighbors. Additionally, in graphs exhibiting
homophily—where connected nodes tend to share similar attributes—sensitive information
can be inferred from neighboring nodes. The use of GNNs for graph encoding further
exacerbates these risks. The message-passing mechanism in GNNs propagates features across
nodes and edges, potentially leading to the leakage of sensitive or confidential information
during the propagation process [534]. Addressing these challenges requires advanced privacypreserving techniques that account for the interconnected nature of graphs, rather than treating
the data as independent and identically distributed (iid). Privacy protection must be designed
across the entire GraphRAG pipeline, from graph construction to generation, ensuring that
sensitive information remains secure while preserving the graph’s utility for downstream tasks.
**– Explainability: Explainability is essential for fostering trust in RAG systems, especially**
in high-stakes domains like law, healthcare, and finance [550, 31, 500, 99]. Compared to
traditional RAG, GraphRAG offers enhanced explainability through the explicit relationships
encoded between nodes in the graph [2]. These explicit connections allow the system to
generate clear and interpretable explanations tailored to specific tasks, making it more transparent and trustworthy for end users. For example, in the mutli-hop question answering, the
GraphRAG can be leveraged to generate reasoning paths to the question [272, 406, 381, 185].
These paths provide step-by-step explanations, allowing users to understand how intermediate
conclusions were derived and verify the relevance and correctness of the system’s logic. Similarly, in molecular property prediction, specific subgraphs representing functional groups or
structural motifs can be utilized to explain predictions, linking molecular features to observed
properties or behaviors. However, achieving such explainability requires GraphRAG to retrieve
reasonable and contextually relevant subgraphs, which remains a challenge. Ensuring that
explanations remain faithful to the underlying model logic while balancing comprehensiveness
and simplicity will be critical challenges for future research.
**10.6** **Evaluation of GraphRAG.**
Evaluating the performance of a GraphRAG system is challenging due to its complex, multicomponent nature. Standard benchmarks may not fully capture the nuances of graph-based
construction, retrieval, organization, and generation, so tailored benchmarks are essential to assess
each component and their overall impact on the system.
- Component-Level Optimality: Each component’s performance directly influences the GraphRAG
system as a whole. Evaluating each component, such as Graph Construction, Retriever, Organizer,
and Generator, requires specific designs suited to their unique roles. This may involve constructing
different datasets and evaluation metrics that align with the intended function of each component.
-----
- End-to-End Benchmarks: To assess the system’s overall effectiveness, comprehensive end-to-end
benchmarks are essential. These should evaluate the quality of generated outputs, system response
time, efficiency, and resource utilization, providing a holistic view of GraphRAG’s performance in
real-world applications.
- Task and Domain-Specific Evaluation: Different tasks and domains may impose unique requirements on the GraphRAG system, necessitating specialized designs for each component. Methods
that perform well in one domain may not generalize effectively to others, highlighting the need for
diverse benchmarks. Additionally, Multi-task and multi-domain evaluations help determine the
system’s adaptability and effectiveness across varied contexts.
- Trustworthiness benchmark. Ensuring the trustworthiness of the GraphRAG system is critical,
especially in applications where decisions rely heavily on accurate and unbiased information. A
trustworthiness benchmark should evaluate aspects such as robustness against adversarial inputs,
data privacy protection, and transparency in how answers are derived, ensuring that outputs are
explainable and reliable.
**10.7** **New Applications**
While we have introduced applications of GraphRAG in several domains and tasks, there are still many
domains that can leverage GraphRAG, such as Code generation [262] and robust cyber defense [330].
Extending its use to other domains is promising but presents unique challenges. Determining how to
adapt GraphRAG effectively for new fields requires understanding the specific requirements, data
structures, and graph configurations unique to each application area. Developing tailored strategies
to construct, retrieve, organize, and generate data for diverse domains is essential for maximizing
GraphRAG’s adaptability and effectiveness.
### 11 Conclusion
In this survey, we first introduced the demand and rationale behind GraphRAG, highlighting its
ability to enhance retrieval-augmented generation by integrating graph-structured information. We
then unified the architectural designs of existing GraphRAG approaches into a holistic framework,
comprising five key components: graph construction, retriever, organizer, generator, and data source.
For each component, we reviewed representative techniques. Given the diversity of graph structures
and their applications across different domains, we also explored GraphRAG designs tailored to
specific domains. By reviewing GraphRAG applications across varied domains—from knowledge
graphs and document graphs to scientific and social graphs, we illustrated how its flexibility allows
it to meet unique demands and address a wide range of tasks. Finally we discuss challenges and
opportunities that have the potential to push the boundaries for GraphRAG.
### References
[1] Josh Abramson, Jonas Adler, Jack Dunger, Richard Evans, Tim Green, Alexander Pritzel, Olaf
Ronneberger, Lindsay Willmore, Andrew J Ballard, Joshua Bambrick, et al. Accurate structure
prediction of biomolecular interactions with alphafold 3. Nature, pages 1–3, 2024.
[2] Hasan Abu-Rasheed, Mohamad Hussam Abdulsalam, Christian Weber, and Madjid Fathi.
Supporting student decisions on learning recommendations: An llm-based chatbot with knowledge graph contextualization for conversational explainability and mentoring. arXiv preprint
_arXiv:2401.08517, 2024._
[3] Oshin Agarwal, Heming Ge, Siamak Shakeri, and Rami Al-Rfou. Knowledge graph based
synthetic corpus generation for knowledge-enhanced language model pre-training. arXiv
_preprint arXiv:2010.12688, 2020._
[4] Charu C Aggarwal, Yao Li, S Yu Philip, and Yuchen Zhao. On edge classification in networks
with structure and content. In 2017 IEEE 33rd international conference on data engineering
_(ICDE), pages 187–190. IEEE, 2017._
[5] Nesreen K Ahmed, Ryan Rossi, John Boaz Lee, Theodore L Willke, Rong Zhou, Xiangnan Kong, and Hoda Eldardiry. Learning role-based graph embeddings. arXiv preprint
_arXiv:1802.02896, 2018._
-----
[6] Tareq Al-Moslmi, Marc Gallofré Ocaña, Andreas L Opdahl, and Csaba Veres. Named entity
extraction for knowledge graphs: A literature overview. IEEE Access, 8:32862–32881, 2020.
[7] Réka Albert and Albert-László Barabási. Statistical mechanics of complex networks. Reviews
_of modern physics, 74(1):47, 2002._
[8] Ralitsa Angelova and Gerhard Weikum. Graph-based text classification: learn from your
neighbors. In Proceedings of the 29th annual international ACM SIGIR conference on Research
_and development in information retrieval, pages 485–492, 2006._
[9] Vlad Argatu, Aaron Haag, and Oliver Lohse. Joint embeddings for graph instruction tuning.
_arXiv preprint arXiv:2405.20684, 2024._
[10] Akari Asai, Kazuma Hashimoto, Hannaneh Hajishirzi, Richard Socher, and Caiming Xiong.
Learning to retrieve reasoning paths over wikipedia graph for question answering. arXiv
_preprint arXiv:1911.10470, 2019._
[11] Akari Asai, Sewon Min, Zexuan Zhong, and Danqi Chen. Retrieval-based language models and
applications. In Proceedings of the 61st Annual Meeting of the Association for Computational
_Linguistics (Volume 6: Tutorial Abstracts), pages 41–46, 2023._
[12] Hiteshwar Kumar Azad and Akshay Deepak. Query expansion techniques for information
retrieval: a survey. Information Processing & Management, 56(5):1698–1735, 2019.
[13] Gregor Bachmann and Vaishnavh Nagarajan. The pitfalls of next-token prediction. arXiv
_preprint arXiv:2403.06963, 2024._
[14] Jinheon Baek, Alham Fikri Aji, and Amir Saffari. Knowledge-augmented language
model prompting for zero-shot knowledge graph question answering. _arXiv preprint_
_arXiv:2306.04136, 2023._
[15] Jinze Bai, Jialin Wang, Zhao Li, Donghui Ding, Ji Zhang, and Jun Gao. Atj-net: Autotable-join network for automatic learning on relational databases. In Proceedings of the Web
_Conference 2021, pages 1540–1551, 2021._
[16] Song Bai, Feihu Zhang, and Philip HS Torr. Hypergraph convolution and hypergraph attention.
_Pattern Recognition, 110:107637, 2021._
[17] Laura Banarescu, Claire Bonial, Shu Cai, Madalina Georgescu, Kira Griffitt, Ulf Hermjakob,
Kevin Knight, Philipp Koehn, Martha Palmer, and Nathan Schneider. Abstract meaning
representation for sembanking. In Proceedings of the 7th linguistic annotation workshop and
_interoperability with discourse, pages 178–186, 2013._
[18] Albert-László Barabási and Réka Albert. Emergence of scaling in random networks. science,
286(5439):509–512, 1999.
[19] Aseem Baranwal, Kimon Fountoulakis, and Aukosh Jagannath. Graph convolution for semisupervised classification: Improved linear separability and out-of-distribution generalization.
_arXiv preprint arXiv:2102.06966, 2021._
[20] Lina Bariah, Qiyang Zhao, Hang Zou, Yu Tian, Faouzi Bader, and Merouane Debbah. Large
generative ai models for telecom: The next big thing? IEEE Communications Magazine, 2024.
[21] Scott Barnett, Stefanus Kurniawan, Srikanth Thudumu, Zach Brannelly, and Mohamed Abdelrazek. Seven failure points when engineering a retrieval augmented generation system.
In Proceedings of the IEEE/ACM 3rd International Conference on AI Engineering-Software
_Engineering for AI, pages 194–199, 2024._
[22] Alev Baysoy, Zhiliang Bai, Rahul Satija, and Rong Fan. The technological landscape and
applications of single-cell multi-omics. Nature Reviews Molecular Cell Biology, 24(10):
695–713, 2023.
[23] Gleb Bazhenov, Oleg Platonov, and Liudmila Prokhorenkova. Tabgraphs: A benchmark and
strong baselines for learning on graphs with tabular node features. arXiv e-prints, pages
arXiv–2409, 2024.
-----
[24] Maciej Besta, Nils Blach, Ales Kubicek, Robert Gerstenberger, Michal Podstawski, Lukas
Gianinazzi, Joanna Gajda, Tomasz Lehmann, Hubert Niewiadomski, Piotr Nyczyk, et al.
Graph of thoughts: Solving elaborate problems with large language models. In Proceedings of
_the AAAI Conference on Artificial Intelligence, volume 38, pages 17682–17690, 2024._
[25] Olivier Bodenreider. The unified medical language system (umls): integrating biomedical
terminology. Nucleic acids research, 32(suppl_1):D267–D270, 2004.
[26] Cristian Bodnar, Fabrizio Frasca, Nina Otter, Yuguang Wang, Pietro Lio, Guido F Montufar,
and Michael Bronstein. Weisfeiler and lehman go cellular: Cw networks. Advances in neural
_information processing systems, 34:2625–2640, 2021._
[27] Kurt Bollacker, Colin Evans, Praveen Paritosh, Tim Sturge, and Jamie Taylor. Freebase: a
collaboratively created graph database for structuring human knowledge. In Proceedings of
_the 2008 ACM SIGMOD international conference on Management of data, pages 1247–1250,_
2008.
[28] Stephen Bonner, Ian P Barrett, Cheng Ye, Rowan Swiers, Ola Engkvist, Andreas Bender,
Charles Tapley Hoyt, and William L Hamilton. A review of biomedical datasets relating to
drug discovery: a knowledge graph perspective. Briefings in Bioinformatics, 23(6):bbac404,
2022.
[29] Jason D Buenrostro, Paul G Giresi, Lisa C Zaba, Howard Y Chang, and William J Greenleaf.
Transposition of native chromatin for multimodal regulatory analysis and personal epigenomics.
_Nature methods, 10(12):1213, 2013._
[30] Odma Byambasuren, Yunfei Yang, Zhifang Sui, Damai Dai, Baobao Chang, Sujian Li, and
Hongying Zan. Preliminary study on the construction of chinese medical knowledge graph.
_Journal of Chinese Information Processing, 33(10):1–9, 2019._
[31] Erik Cambria, Lorenzo Malandri, Fabio Mercorio, Navid Nobani, and Andrea Seveso. Xai
meets llms: A survey of the relation between explainable ai and large language models. arXiv
_preprint arXiv:2407.15248, 2024._
[32] Antonio Candelieri, Dante Conti, and Francesco Archetti. A graph based analysis of leak
localization in urban water networks. Procedia Engineering, 70:228–237, 2014.
[33] Chris Cannings. Random geometric graphs, 2005.
[34] Riccardo Cappuzzo, Saravanan Thirumuruganathan, and Paolo Papotti. Relational data
imputation with graph neural networks. In EDBT/ICDT 2024, 27th International Conference
_on Extending Database Technology, 2024._
[35] Jaime Carbonell and Jade Goldstein. The use of mmr, diversity-based reranking for reordering
documents and producing summaries. In Proceedings of the 21st annual international ACM
_SIGIR conference on Research and development in information retrieval, pages 335–336,_
1998.
[36] Ziwei Chai, Tianjie Zhang, Liang Wu, Kaiqiao Han, Xiaohai Hu, Xuanwen Huang, and Yang
Yang. Graphllm: Boosting graph reasoning ability of large language model. arXiv preprint
_arXiv:2310.05845, 2023._
[37] Payal Chandak, Kexin Huang, and Marinka Zitnik. Building a knowledge graph to enable
precision medicine. Scientific Data, 10(1):67, 2023.
[38] Heng Chang, Jiangnan Ye, Alejo Lopez-Avila, Jinhua Du, and Jia Li. Path-based explanation
for knowledge graph completion. In Proceedings of the 30th ACM SIGKDD Conference on
_Knowledge Discovery and Data Mining, pages 231–242, 2024._
[39] April Chen, Ryan A Rossi, Namyong Park, Puja Trivedi, Yu Wang, Tong Yu, Sungchul Kim,
Franck Dernoncourt, and Nesreen K Ahmed. Fairness-aware graph neural networks: A survey.
_ACM Transactions on Knowledge Discovery from Data, 18(6):1–23, 2024._
-----
[40] Bohan Chen and Andrea L Bertozzi. Autokg: Efficient automated knowledge graph generation
for language models. In 2023 IEEE International Conference on Big Data (BigData), pages
3117–3126. IEEE, 2023.
[41] Chaochao Chen, Fei Zheng, Jamie Cui, Yuwei Cao, Guanfeng Liu, Jia Wu, and Jun Zhou. Survey and open problems in privacy-preserving knowledge graph: merging, query, representation,
completion, and applications. International Journal of Machine Learning and Cybernetics,
pages 1–20, 2024.
[42] Hung-Ting Chen, Michael JQ Zhang, and Eunsol Choi. Rich knowledge sources bring complex
knowledge conflicts: Recalibrating models to reflect conflicting evidence. arXiv preprint
_arXiv:2210.13701, 2022._
[43] Hung-Ting Chen, Fangyuan Xu, Shane A Arora, and Eunsol Choi. Understanding retrieval
augmentation for long-form question answering. arXiv preprint arXiv:2310.12150, 2023.
[44] Katrina Chen, Xiuqin Liang, Zheng Ma, and Zhibin Zhang. Gedi: A graph-based end-toend data imputation framework. In 2023 IEEE 35th International Conference on Tools with
_Artificial Intelligence (ICTAI), pages 723–730. IEEE, 2023._
[45] Moye Chen, Wei Li, Jiachen Liu, Xinyan Xiao, Hua Wu, and Haifeng Wang. Sgsum: transforming multi-document summarization into sub-graph selection. arXiv preprint arXiv:2110.12645,
2021.
[46] Pei Chen, Soumajyoti Sarkar, Leonard Lausen, Balasubramaniam Srinivasan, Sheng Zha,
Ruihong Huang, and George Karypis. Hytrel: Hypergraph-enhanced tabular data representation
learning. Advances in Neural Information Processing Systems, 36, 2024.
[47] Penghe Chen, Yu Lu, Vincent W Zheng, Xiyang Chen, and Boda Yang. Knowedu: A system
to construct knowledge graph for education. Ieee Access, 6:31553–31563, 2018.
[48] Qianglong Chen, Feng Ji, Haiqing Chen, and Yin Zhang. Improving commonsense question
answering by graph-based iterative retrieval over multiple knowledge sources. arXiv preprint
_arXiv:2011.02705, 2020._
[49] Runjin Chen, Tong Zhao, Ajay Jaiswal, Neil Shah, and Zhangyang Wang. Llaga: Large
language and graph assistant. arXiv preprint arXiv:2402.08170, 2024.
[50] Tianqi Chen and Carlos Guestrin. Xgboost: A scalable tree boosting system. In Proceedings
_of the 22nd acm sigkdd international conference on knowledge discovery and data mining,_
pages 785–794, 2016.
[51] Weize Chen, Yusheng Su, Jingwei Zuo, Cheng Yang, Chenfei Yuan, Chen Qian, Chi-Min Chan,
Yujia Qin, Yaxi Lu, Ruobing Xie, et al. Agentverse: Facilitating multi-agent collaboration and
exploring emergent behaviors in agents. arXiv preprint arXiv:2308.10848, 2(4):6, 2023.
[52] Xiaohui Chen, Yinkai Wang, Jiaxing He, Yuanqi Du, Soha Hassoun, Xiaolin Xu, and Li-Ping
Liu. Graph generative pre-trained transformer. arXiv preprint arXiv:2501.01073, 2025.
[53] Xiaojun Chen, Shengbin Jia, and Yang Xiang. A review: Knowledge reasoning over knowledge
graph. Expert systems with applications, 141:112948, 2020.
[54] Xinran Chen, Xuanang Chen, Ben He, Tengfei Wen, and Le Sun. Analyze, generate and refine:
Query expansion with llms for zero-shot open-domain qa. In Findings of the Association for
_Computational Linguistics ACL 2024, pages 11908–11922, 2024._
[55] Zhikai Chen, Haitao Mao, Hongzhi Wen, Haoyu Han, Wei Jin, Haiyang Zhang, Hui Liu, and
Jiliang Tang. Label-free node classification on graphs with large language models (llms). arXiv
_preprint arXiv:2310.04668, 2023._
[56] Zhikai Chen, Haitao Mao, Hang Li, Wei Jin, Hongzhi Wen, Xiaochi Wei, Shuaiqiang Wang,
Dawei Yin, Wenqi Fan, Hui Liu, et al. Exploring the potential of large language models (llms)
in learning on graphs. ACM SIGKDD Explorations Newsletter, 25(2):42–61, 2024.
-----
[57] Dawei Cheng, Yujia Ye, Sheng Xiang, Zhenwei Ma, Ying Zhang, and Changjun Jiang. Antimoney laundering by group-aware deep graph learning. IEEE Transactions on Knowledge and
_Data Engineering, 35(12):12444–12457, 2023._
[58] Kewei Cheng, Nesreen K Ahmed, Theodore Willke, and Yizhou Sun. Structure guided prompt:
Instructing large language model in multi-step reasoning by exploring graph structure of the
text. arXiv preprint arXiv:2402.13415, 2024.
[59] Keyuan Cheng, Gang Lin, Haoyang Fei, Lu Yu, Muhammad Asif Ali, Lijie Hu, Di Wang,
et al. Multi-hop question answering under temporal knowledge editing. arXiv preprint
_arXiv:2404.00492, 2024._
[60] Yi Cheng and Xiuli Ma. scgac: a graph attentional architecture for clustering single-cell
rna-seq data. Bioinformatics, 38(8):2187–2193, 2022.
[61] Zhoujun Cheng, Haoyu Dong, Zhiruo Wang, Ran Jia, Jiaqi Guo, Yan Gao, Shi Han, Jian-Guang
Lou, and Dongmei Zhang. Hitab: A hierarchical table dataset for question answering and
natural language generation. arXiv preprint arXiv:2108.06712, 2021.
[62] Nurendra Choudhary and Chandan K Reddy. Complex logical reasoning over knowledge
graphs using large language models. arXiv preprint arXiv:2305.01157, 2023.
[63] Fenia Christopoulou, Makoto Miwa, and Sophia Ananiadou. Connecting the dots: Documentlevel neural relation extraction with edge-oriented graphs. arXiv preprint arXiv:1909.00228,
2019.
[64] Madalina Ciortan and Matthieu Defrance. Gnn-based embedding for clustering scrna-seq data.
_Bioinformatics, 38(4):1037–1044, 2022._
[65] Karl Cobbe, Vineet Kosaraju, Mohammad Bavarian, Mark Chen, Heewoo Jun, Lukasz Kaiser,
Matthias Plappert, Jerry Tworek, Jacob Hilton, Reiichiro Nakano, et al. Training verifiers to
solve math word problems. arXiv preprint arXiv:2110.14168, 2021.
[66] Edgar F Codd. Relational database: A practical foundation for productivity. In ACM Turing
_award lectures, page 1981. 2007._
[67] Luca Cosmo, Anees Kazi, Seyed-Ahmad Ahmadi, Nassir Navab, and Michael Bronstein.
Latent-graph learning for disease prediction. In Medical Image Computing and Computer
_Assisted Intervention–MICCAI 2020: 23rd International Conference, Lima, Peru, October 4–8,_
_2020, Proceedings, Part II 23, pages 643–653. Springer, 2020._
[68] Hejie Cui, Zijie Lu, Pan Li, and Carl Yang. On positional and structural node features for
graph neural networks on non-attributed graphs. In Proceedings of the 31st ACM International
_Conference on Information & Knowledge Management, pages 3898–3902, 2022._
[69] Wanqing Cui, Keping Bi, Jiafeng Guo, and Xueqi Cheng. More: Multi-modal retrieval
augmented generative commonsense reasoning. arXiv preprint arXiv:2402.13625, 2024.
[70] Alexis Cvetkov-Iliev, Alexandre Allauzen, and Gaël Varoquaux. Relational data embeddings
for feature enrichment with background information. Machine Learning, 112(2):687–720,
2023.
[71] Milan Cvitkovic. Supervised learning on relational databases with graph neural networks.
_arXiv preprint arXiv:2002.02046, 2020._
[72] Xinbang Dai, Yuncheng Hua, Tongtong Wu, Yang Sheng, and Guilin Qi. Counter-intuitive:
Large language models can better understand knowledge graphs than we thought. arXiv
_preprint arXiv:2402.11541, 2024._
[73] Xinnan Dai, Qihao Wen, Yifei Shen, Hongzhi Wen, Dongsheng Li, Jiliang Tang, and Caihua
Shan. Revisiting the graph reasoning ability of large language models: Case studies in
translation, connectivity and shortest path. arXiv preprint arXiv:2408.09529, 2024.
[74] Badhan Chandra Das, M Hadi Amini, and Yanzhao Wu. Security and privacy challenges of
large language models: A survey. arXiv preprint arXiv:2402.00888, 2024.
-----
[75] Rajarshi Das, Shehzaad Dhuliawala, Manzil Zaheer, Luke Vilnis, Ishan Durugkar, Akshay
Krishnamurthy, Alex Smola, and Andrew McCallum. Go for a walk and arrive at the answer:
Reasoning over paths in knowledge bases using reinforcement learning. In International
_Conference on Learning Representations, 2018._
[76] Nicola De Cao, Wilker Aziz, and Ivan Titov. Question answering by reasoning across
documents with graph convolutional networks. arXiv preprint arXiv:1808.09920, 2018.
[77] Yashar Deldjoo, Zhankui He, Julian McAuley, Anton Korikov, Scott Sanner, Arnau Ramisa,
René Vidal, Maheswaran Sathiamoorthy, Atoosa Kasirzadeh, and Silvia Milano. A review
of modern recommender systems using generative models (gen-recsys). _arXiv preprint_
_arXiv:2404.00579, 2024._
[78] Julien Delile, Srayanta Mukherjee, Anton Van Pamel, and Leonid Zhukov. Graph-based
retriever captures the long tail of biomedical knowledge. arXiv preprint arXiv:2402.12352,
2024.
[79] Gelei Deng, Yi Liu, Kailong Wang, Yuekang Li, Tianwei Zhang, and Yang Liu. Pandora:
[Jailbreak gpts by retrieval augmented generation poisoning, 2024. URL https://arxiv.](https://arxiv.org/abs/2402.08416)
```
org/abs/2402.08416.
```
[80] Yash Deshpande, Subhabrata Sen, Andrea Montanari, and Elchanan Mossel. Contextual
stochastic block models. Advances in Neural Information Processing Systems, 31, 2018.
[81] Jacob Devlin. Bert: Pre-training of deep bidirectional transformers for language understanding.
_arXiv preprint arXiv:1810.04805, 2018._
[82] Bhuwan Dhingra, Qiao Jin, Zhilin Yang, William W Cohen, and Ruslan Salakhutdinov.
Neural models for reasoning over multiple mentions using coreference. _arXiv preprint_
_arXiv:1804.05922, 2018._
[83] Laura Dietz, Hannah Bast, Shubham Chatterjee, Jeffrey Dalton, Jian-Yun Nie, and Rodrigo
Nogueira. Neuro-symbolic representations for information retrieval. In Proceedings of the
_46th International ACM SIGIR Conference on Research and Development in Information_
_Retrieval, pages 3436–3439, 2023._
[84] Jiayuan Ding, Lingxiao Li, Qiaolin Lu, Julian Venegas, Yixin Wang, Lidan Wu, Wei Jin,
Hongzhi Wen, Renming Liu, Wenzhuo Tang, et al. Spatialctd: A large-scale tumor microenvironment spatial transcriptomic dataset to evaluate cell type deconvolution for immunooncology. Journal of Computational Biology, 2024.
[85] Jiayuan Ding, Renming Liu, Hongzhi Wen, Wenzhuo Tang, Zhaoheng Li, Julian Venegas,
Runze Su, Dylan Molho, Wei Jin, Yixin Wang, et al. Dance: A deep learning library and
benchmark platform for single-cell analysis. Genome Biology, 25(1):72, 2024.
[86] Kaize Ding, Zhe Xu, Hanghang Tong, and Huan Liu. Data augmentation for deep graph
learning: A survey. ACM SIGKDD Explorations Newsletter, 24(2):61–77, 2022.
[87] Yujuan Ding, Wenqi Fan, Liangbo Ning, Shijie Wang, Hengyun Li, Dawei Yin, Tat-Seng Chua,
and Qing Li. A survey on rag meets llms: Towards retrieval-augmented large language models.
_arXiv preprint arXiv:2405.06211, 2024._
[88] Shibhansh Dohare, Harish Karnick, and Vivek Gupta. Text summarization using abstract
meaning representation. arXiv preprint arXiv:1706.01678, 2017.
[89] Haoyu Dong and Zhiruo Wang. Large language models for tabular data: Progresses and future
directions. In Proceedings of the 47th International ACM SIGIR Conference on Research and
_Development in Information Retrieval, pages 2997–3000, 2024._
[90] Jialin Dong, Bahare Fatemi, Bryan Perozzi, Lin F Yang, and Anton Tsitsulin. Don’t forget to
connect! improving rag with graph-based reranking. arXiv preprint arXiv:2405.18414, 2024.
[91] Kangning Dong and Shihua Zhang. Deciphering spatial domains from spatially resolved
transcriptomics with an adaptive graph attention auto-encoder. Nature communications, 13(1):
1739, 2022.
-----
[92] Li Dong, Shaohan Huang, Furu Wei, Mirella Lapata, Ming Zhou, and Ke Xu. Learning
to generate product reviews from attributes. In Proceedings of the 15th Conference of the
_European Chapter of the Association for Computational Linguistics: Volume 1, Long Papers,_
pages 623–632, 2017.
[93] Claire Donnat, Marinka Zitnik, David Hallac, and Jure Leskovec. Learning structural node
embeddings via diffusion wavelets. In Proceedings of the 24th ACM SIGKDD international
_conference on knowledge discovery & data mining, pages 1320–1329, 2018._
[94] Kounianhua Du, Weinan Zhang, Ruiwen Zhou, Yangkun Wang, Xilong Zhao, Jiarui Jin, Quan
Gan, Zheng Zhang, and David P Wipf. Learning enhanced representation for tabular data
via neighborhood propagation. Advances in Neural Information Processing Systems, 35:
16373–16384, 2022.
[95] Yingpeng Du, Ziyan Wang, Zhu Sun, Haoyan Chua, Hongzhi Liu, Zhonghai Wu, Yining Ma,
Jie Zhang, and Youchen Sun. Large language model with graph convolution for recommendation. arXiv preprint arXiv:2402.08859, 2024.
[96] Dheeru Dua, Casey Graff, et al. Uci machine learning repository. 2017.
[97] Nouha Dziri, Ximing Lu, Melanie Sclar, Xiang Lorraine Li, Liwei Jiang, Bill Yuchen Lin,
Sean Welleck, Peter West, Chandra Bhagavatula, Ronan Le Bras, et al. Faith and fate: Limits
of transformers on compositionality. Advances in Neural Information Processing Systems, 36,
2024.
[98] Darren Edge, Ha Trinh, Newman Cheng, Joshua Bradley, Alex Chao, Apurva Mody, Steven
Truitt, and Jonathan Larson. From local to global: A graph rag approach to query-focused
summarization. arXiv preprint arXiv:2404.16130, 2024.
[99] Jonas Elsborg and Marco Salvatore. Using llms and explainable ml to analyze biomarkers at
single-cell level for improved understanding of diseases. Biomolecules, 13(10):1516, 2023.
[100] P ERDdS and A R&wi. On random graphs i. Publ. math. debrecen, 6(290-297):18, 1959.
[101] Nick Erickson, Jonas Mueller, Alexander Shirkov, Hang Zhang, Pedro Larroy, Mu Li, and
Alexander Smola. Autogluon-tabular: Robust and accurate automl for structured data. arXiv
_preprint arXiv:2003.06505, 2020._
[102] Federico Errica. On class distributions induced by nearest neighbor graphs for node classification of tabular data. Advances in Neural Information Processing Systems, 36, 2024.
[103] Peter Ertl. An algorithm to identify functional groups in organic molecules. Journal of
_cheminformatics, 9:1–7, 2017._
[104] Wenqi Fan, Yujuan Ding, Liangbo Ning, Shijie Wang, Hengyun Li, Dawei Yin, Tat-Seng Chua,
and Qing Li. A survey on rag meeting llms: Towards retrieval-augmented large language
models. In Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and
_Data Mining, pages 6491–6501, 2024._
[105] Feiteng Fang, Yuelin Bai, Shiwen Ni, Min Yang, Xiaojun Chen, and Ruifeng Xu. Enhancing
noise robustness of retrieval-augmented language models with adaptive adversarial training.
_arXiv preprint arXiv:2405.20978, 2024._
[106] Jinyuan Fang, Zaiqiao Meng, and Craig Macdonald. Reano: Optimising retrieval-augmented
reader models through knowledge graph generation. In Proceedings of the 62nd Annual
_Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages_
2094–2112, 2024.
[107] Xi Fang, Weijie Xu, Fiona Anting Tan, Jiani Zhang, Ziqing Hu, Yanjun Jane Qi, Scott
Nickleach, Diego Socolinsky, Srinivasan Sengamedu, Christos Faloutsos, et al. Large language
models (llms) on tabular data: Prediction, generation, and understanding-a survey. 2024.
[108] Yuwei Fang, Siqi Sun, Zhe Gan, Rohit Pillai, Shuohang Wang, and Jingjing Liu. Hierarchical
graph network for multi-hop question answering. arXiv preprint arXiv:1911.03631, 2019.
-----
[109] Bahare Fatemi, Jonathan Halcrow, and Bryan Perozzi. Talk like a graph: Encoding graphs for
large language models. arXiv preprint arXiv:2310.04560, 2023.
[110] Yanlin Feng, Xinyue Chen, Bill Yuchen Lin, Peifeng Wang, Jun Yan, and Xiang Ren. Scalable
multi-hop relational reasoning for knowledge-aware question answering. In Proceedings of
_the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages_
1295–1309, 2020.
[111] Yifan Feng, Haoxuan You, Zizhao Zhang, Rongrong Ji, and Yue Gao. Hypergraph neural
networks. In Proceedings of the AAAI conference on artificial intelligence, volume 33, pages
3558–3565, 2019.
[112] Paolo Ferragina and Ugo Scaiella. Tagme: on-the-fly annotation of short text fragments (by
wikipedia entities). In Proceedings of the 19th ACM international conference on Information
_and knowledge management, pages 1625–1628, 2010._
[113] Miquel Ferriol-Galmés, Jordi Paillisse, José Suárez-Varela, Krzysztof Rusek, Shihan Xiao,
Xiang Shi, Xiangle Cheng, Pere Barlet-Ros, and Albert Cabellos-Aparicio. Routenet-fermi:
Network modeling with graph neural networks. IEEE/ACM transactions on networking, 31(6):
3080–3095, 2023.
[114] Matthias Fey and Jan Eric Lenssen. Fast graph representation learning with pytorch geometric.
_arXiv preprint arXiv:1903.02428, 2019._
[115] Matthias Fey, Weihua Hu, Kexin Huang, Jan Eric Lenssen, Rishabh Ranjan, Joshua Robinson,
Rex Ying, Jiaxuan You, and Jure Leskovec. Relational deep learning: Graph representation
learning on relational databases. arXiv preprint arXiv:2312.04615, 2023.
[116] Luke Friedman, Sameer Ahuja, David Allen, Zhenning Tan, Hakim Sidahmed, Changbo Long,
Jun Xie, Gabriel Schubiner, Ajay Patel, Harsh Lara, et al. Leveraging large language models
in conversational recommender systems. arXiv preprint arXiv:2305.07961, 2023.
[117] Yanglan Gan, Xingyu Huang, Guobing Zou, Shuigeng Zhou, and Jihong Guan. Deep structural
clustering for single-cell rna-seq data jointly through autoencoder and graph neural network.
_Briefings in Bioinformatics, 23(2):bbac018, 2022._
[118] Hanning Gao, Lingfei Wu, Po Hu, Zhihua Wei, Fangli Xu, and Bo Long. Graph-augmented
learning to rank for querying large-scale knowledge graph. In Proceedings of the 2nd Confer_ence of the Asia-Pacific Chapter of the Association for Computational Linguistics and the 12th_
_International Joint Conference on Natural Language Processing (Volume 1: Long Papers),_
pages 82–92, 2022.
[119] Yifu Gao, Linbo Qiao, Zhigang Kan, Zhihua Wen, Yongquan He, and Dongsheng Li. Twostage generative question answering on temporal knowledge graph using large language models.
_arXiv preprint arXiv:2402.16568, 2024._
[120] Yunfan Gao, Yun Xiong, Xinyu Gao, Kangxiang Jia, Jinliu Pan, Yuxi Bi, Yi Dai, Jiawei Sun,
and Haofen Wang. Retrieval-augmented generation for large language models: A survey.
_arXiv preprint arXiv:2312.10997, 2023._
[121] Anna Gaulton, Louisa J Bellis, A Patricia Bento, Jon Chambers, Mark Davies, Anne Hersey,
Yvonne Light, Shaun McGlinchey, David Michalovich, Bissan Al-Lazikani, et al. Chembl: a
large-scale bioactivity database for drug discovery. Nucleic acids research, 40(D1):D1100–
D1107, 2012.
[122] Yuyao Ge, Shenghua Liu, Lingrui Mei, Lizhe Chen, and Xueqi Cheng. Can graph descriptive
order affect solving graph problems with llms? arXiv preprint arXiv:2402.07140v4, 2024.
[123] Goran Glavaš and Jan Šnajder. Event-centered information retrieval using kernels on event
graphs. In Proceedings of TextGraphs-8 graph-based methods for natural language processing,
pages 1–5, 2013.
-----
[124] Adam Goodge, Bryan Hooi, See-Kiong Ng, and Wee Siong Ng. Lunar: Unifying local outlier
detection methods via graph neural networks. In Proceedings of the AAAI Conference on
_Artificial Intelligence, volume 36, pages 6737–6745, 2022._
[125] Yury Gorishniy, Ivan Rubachev, Valentin Khrulkov, and Artem Babenko. Revisiting deep
learning models for tabular data. Advances in Neural Information Processing Systems, 34:
18932–18943, 2021.
[126] Léo Grinsztajn, Edouard Oyallon, and Gaël Varoquaux. Why do tree-based models still
outperform deep learning on typical tabular data? Advances in neural information processing
_systems, 35:507–520, 2022._
[127] Aditya Grover and Jure Leskovec. node2vec: Scalable feature learning for networks. In
_Proceedings of the 22nd ACM SIGKDD international conference on Knowledge discovery and_
_data mining, pages 855–864, 2016._
[128] Yifan Gu, Xuebing Yang, Lei Tian, Hongyu Yang, Jicheng Lv, Chao Yang, Jinwei Wang,
Jianing Xi, Guilan Kong, and Wensheng Zhang. Structure-aware siamese graph neural
networks for encounter-level patient similarity learning. Journal of Biomedical Informatics,
127:104027, 2022.
[129] Huifeng Guo, Ruiming Tang, Yunming Ye, Zhenguo Li, and Xiuqiang He. Deepfm:
a factorization-machine based neural network for ctr prediction. _arXiv preprint_
_arXiv:1703.04247, 2017._
[130] Jiayan Guo, Lun Du, Hengyu Liu, Mengyu Zhou, Xinyi He, and Shi Han. Gpt4graph: Can large
language models understand graph structured data? an empirical evaluation and benchmarking.
_arXiv preprint arXiv:2305.15066, 2023._
[131] Qingyu Guo, Fuzhen Zhuang, Chuan Qin, Hengshu Zhu, Xing Xie, Hui Xiong, and Qing He.
A survey on knowledge graph-based recommender systems. IEEE Transactions on Knowledge
_and Data Engineering, 34(8):3549–3568, 2020._
[132] Taicheng Guo, Bozhao Nan, Zhenwen Liang, Zhichun Guo, Nitesh Chawla, Olaf Wiest,
Xiangliang Zhang, et al. What can large language models do in chemistry? a comprehensive
benchmark on eight tasks. Advances in Neural Information Processing Systems, 36:59662–
59688, 2023.
[133] Tiezheng Guo, Qingwen Yang, Chen Wang, Yanyi Liu, Pan Li, Jiawei Tang, Dapeng Li,
and Yingyou Wen. Knowledgenavigator: Leveraging large language models for enhanced
reasoning over knowledge graph. Complex & Intelligent Systems, 10(5):7063–7076, 2024.
[134] Tiezheng Guo, Qingwen Yang, Chen Wang, Yanyi Liu, Pan Li, Jiawei Tang, Dapeng Li,
and Yingyou Wen. Knowledgenavigator: Leveraging large language models for enhanced
reasoning over knowledge graph. Complex & Intelligent Systems, pages 1–14, 2024.
[135] Wei Guo, Rong Su, Renhao Tan, Huifeng Guo, Yingxue Zhang, Zhirong Liu, Ruiming Tang,
and Xiuqiang He. Dual graph enhanced embedding neural network for ctr prediction. In
_Proceedings of the 27th ACM SIGKDD conference on knowledge discovery & data mining,_
pages 496–504, 2021.
[136] Xiawei Guo, Yuhan Quan, Huan Zhao, Quanming Yao, Yong Li, and Weiwei Tu. Tabgnn:
Multiplex graph neural network for tabular data prediction. arXiv preprint arXiv:2108.09127,
2021.
[137] Zirui Guo, Lianghao Xia, Yanhua Yu, Tu Ao, and Chao Huang. Lightrag: Simple and fast
retrieval-augmented generation. arXiv preprint arXiv:2410.05779, 2024.
[138] Will Hamilton, Zhitao Ying, and Jure Leskovec. Inductive representation learning on large
graphs. Advances in neural information processing systems, 30, 2017.
[139] William L Hamilton. Graph representation learning. Morgan & Claypool Publishers, 2020.
-----
[140] Haoyu Han, Juanhui Li, Wei Huang, Xianfeng Tang, Hanqing Lu, Chen Luo, Hui Liu, and
Jiliang Tang. Node-wise filtering in graph neural networks: A mixture of experts approach.
_arXiv preprint arXiv:2406.03464, 2024._
[141] Xu Han, Tianyu Gao, Yankai Lin, Hao Peng, Yaoliang Yang, Chaojun Xiao, Zhiyuan Liu, Peng
Li, Maosong Sun, and Jie Zhou. More data, more relations, more context and more openness:
A review and outlook for relation extraction. arXiv preprint arXiv:2004.03186, 2020.
[142] Shibo Hao, Yi Gu, Haodi Ma, Joshua Jiahua Hong, Zhen Wang, Daisy Zhe Wang, and
Zhiting Hu. Reasoning with language model is planning with world model. arXiv preprint
_arXiv:2305.14992, 2023._
[143] F Maxwell Harper and Joseph A Konstan. The movielens datasets: History and context. Acm
_transactions on interactive intelligent systems (tiis), 5(4):1–19, 2015._
[144] Gaole He, Yunshi Lan, Jing Jiang, Wayne Xin Zhao, and Ji-Rong Wen. Improving multihop knowledge base question answering by learning intermediate supervision signals. In
_Proceedings of the 14th ACM international conference on web search and data mining, pages_
553–561, 2021.
[145] Pengfei He, Zitao Li, Yue Xing, Yaling Li, Jiliang Tang, and Bolin Ding. Make llms better zeroshot reasoners: Structure-orientated autonomous reasoning. arXiv preprint arXiv:2410.19000,
2024.
[146] Ruining He and Julian McAuley. Ups and downs: Modeling the visual evolution of fashion
trends with one-class collaborative filtering. In proceedings of the 25th international conference
_on world wide web, pages 507–517, 2016._
[147] Xiaoxin He, Yijun Tian, Yifei Sun, Nitesh V Chawla, Thomas Laurent, Yann LeCun, Xavier
Bresson, and Bryan Hooi. G-retriever: Retrieval-augmented generation for textual graph
understanding and question answering. arXiv preprint arXiv:2402.07630, 2024.
[148] Zhongmou He, Jing Zhu, Shengyi Qian, Joyce Chai, and Danai Koutra. Linkgpt: Teaching
large language models to predict missing links. arXiv preprint arXiv:2406.04640, 2024.
[149] Nima Hemmati and Gholamreza Ghassem-Sani. Multi-hop question answering using sparse
graphs. Engineering Applications of Artificial Intelligence, 126:107128, 2023.
[150] Paul W Holland, Kathryn Blackmond Laskey, and Samuel Leinhardt. Stochastic blockmodels:
First steps. Social networks, 5(2):109–137, 1983.
[151] Matthew Honnibal and Ines Montani. spacy 2: Natural language understanding with bloom
embeddings, convolutional neural networks and incremental parsing. To appear, 7(1):411–420,
2017.
[152] Yupeng Hou, Junjie Zhang, Zihan Lin, Hongyu Lu, Ruobing Xie, Julian McAuley, and
Wayne Xin Zhao. Large language models are zero-shot rankers for recommender systems. In
_European Conference on Information Retrieval, pages 364–381. Springer, 2024._
[153] Edward J Hu, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, Weizhu
Chen, et al. Lora: Low-rank adaptation of large language models. In International Conference
_on Learning Representations, 2022._
[154] Jian Hu, Xiangjie Li, Kyle Coleman, Amelia Schroeder, Nan Ma, David J Irwin, Edward B Lee,
Russell T Shinohara, and Mingyao Li. Spagcn: Integrating gene expression, spatial location
and histology to identify spatial domains and spatially variable genes by graph convolutional
network. Nature methods, 18(11):1342–1351, 2021.
[155] Linmei Hu, Tianchi Yang, Luhao Zhang, Wanjun Zhong, Duyu Tang, Chuan Shi, Nan Duan,
and Ming Zhou. Compare to the knowledge: Graph neural fake news detection with external
knowledge. In Proceedings of the 59th Annual Meeting of the Association for Computational
_Linguistics and the 11th International Joint Conference on Natural Language Processing_
_(Volume 1: Long Papers), pages 754–763, 2021._
-----
[156] Weihua Hu, Yiwen Yuan, Zecheng Zhang, Akihiro Nitta, Kaidi Cao, Vid Kocijan, Jure
Leskovec, and Matthias Fey. Pytorch frame: A modular framework for multi-modal tabular
learning. arXiv preprint arXiv:2404.00776, 2024.
[157] Ziniu Hu, Yizhou Sun, and Kai-Wei Chang. Relation-guided pre-training for open-domain
question answering. arXiv preprint arXiv:2109.10346, 2021.
[158] Ziniu Hu, Yichong Xu, Wenhao Yu, Shuohang Wang, Ziyi Yang, Chenguang Zhu, Kai-Wei
Chang, and Yizhou Sun. Empowering language models with knowledge graph reasoning
for open-domain question answering. In Proceedings of the 2022 Conference on Empirical
_Methods in Natural Language Processing, pages 9562–9581, 2022._
[159] Chao Huang, Jiahui Chen, Lianghao Xia, Yong Xu, Peng Dai, Yanqing Chen, Liefeng Bo,
Jiashu Zhao, and Jimmy Xiangji Huang. Graph-enhanced multi-task learning of multi-level
transition dynamics for session-based recommendation. In Proceedings of the AAAI conference
_on artificial intelligence, volume 35, pages 4123–4130, 2021._
[160] Jiaxin Huang, Chunyuan Li, Krishan Subudhi, Damien Jose, Shobana Balakrishnan, Weizhu
Chen, Baolin Peng, Jianfeng Gao, and Jiawei Han. Few-shot named entity recognition: An
empirical baseline study. In Proceedings of the 2021 conference on empirical methods in
_natural language processing, pages 10408–10423, 2021._
[161] Kexin Huang. scgnn: scrna-seq dropout imputation via induced hierarchical cell similarity
graph. arXiv preprint arXiv:2008.03322, 2020.
[162] Lifu Huang, Heng Ji, Kyunghyun Cho, and Clare R Voss. Zero-shot transfer learning for event
extraction. arXiv preprint arXiv:1707.01066, 2017.
[163] Xingyue Huang, Miguel Romero, Ismail Ceylan, and Pablo Barceló. A theory of link prediction via relational weisfeiler-leman on knowledge graphs. Advances in Neural Information
_Processing Systems, 36, 2024._
[164] Yongfeng Huang, Yanyang Li, Yichong Xu, Lin Zhang, Ruyi Gan, Jiaxing Zhang, and Liwei
Wang. Mvp-tuning: Multi-view knowledge retrieval with prompt tuning for commonsense
reasoning. In Proceedings of the 61st Annual Meeting of the Association for Computational
_Linguistics (Volume 1: Long Papers), pages 13417–13432, 2023._
[165] Yongjie Huang and Meng Yang. Breadth first reasoning graph for multi-hop question answering.
In Proceedings of the 2021 Conference of the North American Chapter of the Association for
_Computational Linguistics: Human Language Technologies, pages 5810–5821, 2021._
[166] Yue Huang, Lichao Sun, Haoran Wang, Siyuan Wu, Qihui Zhang, Yuan Li, Chujie Gao,
Yixin Huang, Wenhan Lyu, Yixuan Zhang, et al. Position: Trustllm: Trustworthiness in large
language models. In International Conference on Machine Learning, pages 20166–20270.
PMLR, 2024.
[167] Zhilin Huang, Ling Yang, Xiangxin Zhou, Chujun Qin, Yijie Yu, Xiawu Zheng, Zikun Zhou,
Wentao Zhang, Yu Wang, and Wenming Yang. Interaction-based retrieval-augmented diffusion
models for protein-specific 3d molecule generation. In International Conference on Machine
_Learning, 2024._
[168] Mohamed Manzour Hussien, Angie Nataly Melo, Augusto Luis Ballardini, Carlota Salinas
Maldonado, Rubén Izquierdo, and Miguel Ángel Sotelo. Rag-based explainable prediction
of road users behaviors for automated driving using knowledge graphs and large language
models. arXiv preprint arXiv:2405.00449, 2024.
[169] Filip Ilievski, Pedro Szekely, and Bin Zhang. Cskg: The commonsense knowledge graph. In
_The Semantic Web: 18th International Conference, ESWC 2021, Virtual Event, June 6–10,_
_2021, Proceedings 18, pages 680–696. Springer, 2021._
[170] John J Irwin and Brian K Shoichet. Zinc- a free database of commercially available compounds
for virtual screening. Journal of chemical information and modeling, 45(1):177–182, 2005.
-----
[171] Sergei Ivanov and Liudmila Prokhorenkova. Boost then convolve: Gradient boosting meets
graph neural networks. arXiv preprint arXiv:2101.08543, 2021.
[172] Maor Ivgi, Uri Shaham, and Jonathan Berant. Efficient long-text understanding with short-text
models. Transactions of the Association for Computational Linguistics, 11:284–299, 2023.
[173] Rolf Jagerman, Honglei Zhuang, Zhen Qin, Xuanhui Wang, and Michael Bendersky. Query
expansion by prompting large language models. arXiv preprint arXiv:2305.03653, 2023.
[174] Svante Janson, Tomasz Luczak, and Andrzej Rucinski. Random graphs. John Wiley & Sons,
2011.
[175] Minbyul Jeong, Jiwoong Sohn, Mujeen Sung, and Jaewoo Kang. Improving medical reasoning
through retrieval and self-reflection with retrieval-augmented large language models. arXiv
_preprint arXiv:2401.15269, 2024._
[176] Soyeong Jeong, Jinheon Baek, Sukmin Cho, Sung Ju Hwang, and Jong C Park. Adaptive-rag:
Learning to adapt retrieval-augmented large language models through question complexity.
_arXiv preprint arXiv:2403.14403, 2024._
[177] Xingguo Ji and Qingmin Meng. Traffic classification based on graph convolutional network.
In 2020 IEEE International Conference on Advances in Electrical Engineering and Computer
_Applications (AEECA), pages 596–601. IEEE, 2020._
[178] Ran Jia, Haoming Guo, Xiaoyuan Jin, Chao Yan, Lun Du, Xiaojun Ma, Tamara Stankovic,
Marko Lozajic, Goran Zoranovic, Igor Ilic, et al. Getpt: Graph-enhanced general table pretraining with alternate attention network. In Proceedings of the 29th ACM SIGKDD Conference
_on Knowledge Discovery and Data Mining, pages 941–950, 2023._
[179] Haifeng Wu Jian Yang, Xuefeng Li. DeepTables: A Deep Learning Python Package for
Tabular Data. https://github.com/DataCanvasIO/DeepTables, 2022. Version 0.2.x.
[180] Albert Q Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh
Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lucile
Saulnier, et al. Mistral 7b. arXiv preprint arXiv:2310.06825, 2023.
[181] Jinhao Jiang, Kun Zhou, Zican Dong, Keming Ye, Wayne Xin Zhao, and Ji-Rong Wen.
Structgpt: A general framework for large language model to reason over structured data. In
_Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing,_
pages 9237–9251, 2023.
[182] Jinhao Jiang, Kun Zhou, Wayne Xin Zhao, Yang Song, Chen Zhu, Hengshu Zhu, and JiRong Wen. Kg-agent: An efficient autonomous agent framework for complex reasoning over
knowledge graph. arXiv preprint arXiv:2402.11163, 2024.
[183] Julie Jiang and Emilio Ferrara. Social-llm: Modeling user behavior at scale using language
models and social network data. arXiv preprint arXiv:2401.00893, 2023.
[184] Weiwei Jiang and Jiayun Luo. Graph neural network for traffic forecasting: A survey. Expert
_systems with applications, 207:117921, 2022._
[185] Xinke Jiang, Ruizhe Zhang, Yongxin Xu, Rihong Qiu, Yue Fang, Zhiyuan Wang, Jinyi Tang,
Hongxin Ding, Xu Chu, Junfeng Zhao, et al. Hykge: A hypothesis knowledge graph enhanced
framework for accurate and reliable medical llms responses. arXiv preprint arXiv:2312.15883,
2024.
[186] Bowen Jin, Chulin Xie, Jiawei Zhang, Kashob Kumar Roy, Yu Zhang, Suhang Wang, Yu Meng,
and Jiawei Han. Graph chain-of-thought: Augmenting large language models by reasoning on
graphs. arXiv preprint arXiv:2404.07103, 2024.
[187] Qiao Jin, Won Kim, Qingyu Chen, Donald C Comeau, Lana Yeganova, W John Wilbur, and
Zhiyong Lu. Medcpt: Contrastive pre-trained transformers with large-scale pubmed search
logs for zero-shot biomedical information retrieval. Bioinformatics, 39(11):btad651, 2023.
-----
[188] Rihui Jin, Jianan Wang, Wei Tan, Yongrui Chen, Guilin Qi, and Wang Hao. Tabprompt:
Graph-based pre-training and prompting for few-shot table understanding. In Findings of the
_Association for Computational Linguistics: EMNLP 2023, pages 7373–7383, 2023._
[189] Zhuoran Jin, Pengfei Cao, Yubo Chen, Kang Liu, Xiaojian Jiang, Jiexin Xu, Qiuxia Li, and
Jun Zhao. Tug-of-war between knowledge: Exploring and resolving knowledge conflicts in
retrieval-augmented language models. arXiv preprint arXiv:2402.14409, 2024.
[190] Manu Joseph. Pytorch tabular: A framework for deep learning with tabular data, 2021.
[191] Mingxuan Ju, Wenhao Yu, Tong Zhao, Chuxu Zhang, and Yanfang Ye. Grape: Knowledge graph enhanced passage reader for open-domain question answering. arXiv preprint
_arXiv:2210.02933, 2022._
[192] Wei Ju, Yifan Wang, Yifang Qin, Zhengyang Mao, Zhiping Xiao, Junyu Luo, Junwei Yang,
Yiyang Gu, Dongjie Wang, Qingqing Long, et al. Towards graph contrastive learning: A
survey and beyond. arXiv preprint arXiv:2405.11868, 2024.
[193] John Jumper, Richard Evans, Alexander Pritzel, Tim Green, Michael Figurnov, Olaf Ronneberger, Kathryn Tunyasuvunakool, Russ Bates, Augustin Žídek, Anna Potapenko, et al.
Highly accurate protein structure prediction with alphafold. nature, 596(7873):583–589, 2021.
[194] Seokho Kang. K-nearest neighbor learning with graph neural networks. Mathematics, 9(8):
830, 2021.
[195] James Max Kanter and Kalyan Veeramachaneni. Deep feature synthesis: Towards automating
data science endeavors. In 2015 IEEE international conference on data science and advanced
_analytics (DSAA), pages 1–10. IEEE, 2015._
[196] Vladimir Karpukhin, Barlas O˘guz, Sewon Min, Patrick Lewis, Ledell Wu, Sergey Edunov,
Danqi Chen, and Wen-tau Yih. Dense passage retrieval for open-domain question answering.
_arXiv preprint arXiv:2004.04906, 2020._
[197] Anees Kazi, Luca Cosmo, Seyed-Ahmad Ahmadi, Nassir Navab, and Michael M Bronstein.
Differentiable graph module (dgm) for graph convolutional networks. IEEE Transactions on
_Pattern Analysis and Machine Intelligence, 45(2):1606–1617, 2022._
[198] Sara Kemper, Justin Cui, Kai Dicarlantonio, Kathy Lin, Danjie Tang, Anton Korikov, and
Scott Sanner. Retrieval-augmented conversational recommendation with prompt-based semistructured natural language state tracking. In Proceedings of the 47th International ACM
_SIGIR Conference on Research and Development in Information Retrieval, pages 2786–2790,_
2024.
[199] Imed Keraghel, Stanislas Morbieu, and Mohamed Nadif. A survey on recent advances in
named entity recognition. arXiv preprint arXiv:2401.10825, 2024.
[200] Jiho Kim, Yeonsu Kwon, Yohan Jo, and Edward Choi. Kg-gpt: A general framework for
reasoning on knowledge graphs using large language models. In Findings of the Association
_for Computational Linguistics: EMNLP 2023, pages 9410–9421, 2023._
[201] Jiho Kim, Sungjin Park, Yeonsu Kwon, Yohan Jo, James Thorne, and Edward Choi. Factkg:
Fact verification via reasoning on knowledge graphs. arXiv preprint arXiv:2305.06590, 2023.
[202] Jihyeok Kim, Seungtaek Choi, Reinald Kim Amplayo, and Seung-won Hwang. Retrievalaugmented controllable review generation. In Proceedings of the 28th International Conference
_on Computational Linguistics, pages 2284–2295, 2020._
[203] Kiseung Kim and Jay-Yoon Lee. Re-rag: Improving open-domain qa performance and
interpretability with relevance estimator in retrieval-augmented generation. arXiv preprint
_arXiv:2406.05794, 2024._
[204] Sunghwan Kim, Jie Chen, Tiejun Cheng, Asta Gindulyte, Jia He, Siqian He, Qingliang Li,
Benjamin A Shoemaker, Paul A Thiessen, Bo Yu, et al. Pubchem 2019 update: improved
access to chemical data. Nucleic acids research, 47(D1):D1102–D1109, 2019.
-----
[205] Yejin Kim, Eojin Kang, Juae Kim, and H Howie Huang. Causal reasoning in large language
models: A knowledge graph approach. In Causality and Large Models@ NeurIPS 2024.
[206] Thomas N Kipf and Max Welling. Semi-supervised classification with graph convolutional
networks. arXiv preprint arXiv:1609.02907, 2016.
[207] Thomas N Kipf and Max Welling. Semi-supervised classification with graph convolutional
networks. In International Conference on Learning Representations, 2017.
[208] Sungho Ko, Hyunjin Cho, Hyungjoo Chae, Jinyoung Yeo, and Dongha Lee. Evidence-focused
fact summarization for knowledge-augmented zero-shot question answering. arXiv preprint
_arXiv:2403.02966, 2024._
[209] Furkan Kocayusufoglu, Arlei Silva, and Ambuj K Singh. Flowgen: A generative model for
flow graphs. In Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery
_and Data Mining, pages 813–823, 2022._
[210] Aleksandra A Kolodziejczyk, Jong Kyoung Kim, Valentine Svensson, John C Marioni, and
Sarah A Teichmann. The technology and biology of single-cell rna sequencing. Molecular
_cell, 58(4):610–620, 2015._
[211] Rik Koncel-Kedziorski, Dhanush Bekal, Yi Luan, Mirella Lapata, and Hannaneh Hajishirzi. Text generation from knowledge graphs with graph transformers. arXiv preprint
_arXiv:1904.02342, 2019._
[212] Lecheng Kong, Jiarui Feng, Hao Liu, Chengsong Huang, Jiaxin Huang, Yixin Chen, and
Muhan Zhang. Gofa: A generative one-for-all model for joint graph language modeling. arXiv
_preprint arXiv:2407.09709, 2024._
[213] Manikanta Kotaru. Adapting foundation models for operator data analytics. In Proceedings of
_the 22nd ACM Workshop on Hot Topics in Networks, pages 172–179, 2023._
[214] Mahnaz Koupaee and William Yang Wang. Wikihow: A large scale text summarization dataset.
_arXiv preprint arXiv:1810.09305, 2018._
[215] Bhawesh Kumar, Charlie Lu, Gauri Gupta, Anil Palepu, David Bellamy, Ramesh Raskar, and
Andrew Beam. Conformal prediction with large language models for multi-choice question
[answering, 2023. URL https://arxiv.org/abs/2305.18404.](https://arxiv.org/abs/2305.18404)
[216] Deepika Kumawat and Vinesh Jain. Pos tagging approaches: A comparison. International
_Journal of Computer Applications, 118(6), 2015._
[217] Greg Landrum et al. Rdkit: A software suite for cheminformatics, computational chemistry,
and predictive modeling. Greg Landrum, 8(31.10):5281, 2013.
[218] Justine Lebrun, Sylvie Philipp-Foliguet, and Philippe-Henri Gosselin. Image retrieval with
graph kernel on regions. In 2008 19th International Conference on Pattern Recognition, pages
1–4. IEEE, 2008.
[219] Heeyoung Lee, Yves Peirsman, Angel Chang, Nathanael Chambers, Mihai Surdeanu, and Dan
Jurafsky. Stanford’s multi-pass sieve coreference resolution system at the conll-2011 shared
task. In Proceedings of the fifteenth conference on computational natural language learning:
_Shared task, pages 28–34, 2011._
[220] Zhicheng Lee, Zhidian Huang, Zijun Yao, Jinxin Liu, Amy Xin, Lei Hou, and Juanzi Li.
Diakop: Dialogue-based knowledge-oriented programming for neural-symbolic knowledge
base question answering. In Proceedings of the 33rd ACM International Conference on
_Information and Knowledge Management, pages 5234–5238, 2024._
[221] Yibin Lei, Yu Cao, Tianyi Zhou, Tao Shen, and Andrew Yates. Corpus-steered query expansion
with large language models. arXiv preprint arXiv:2402.18031, 2024.
[222] Jure Leskovec, Natasa Milic-Frayling, and Marko Grobelnik. Extracting summary sentences
based on the document semantic graph. 2005.
-----
[223] Cheng-Te Li, Yu-Che Tsai, Chih-Yao Chen, and Jay Chiehen Liao. Graph neural networks for tabular data learning: A survey with taxonomy and directions. arXiv preprint
_arXiv:2401.02143, 2024._
[224] Dawei Li, Shu Yang, Zhen Tan, Jae Young Baik, Sunkwon Yun, Joseph Lee, Aaron Chacko,
Bojian Hou, Duy Duong-Tran, Ying Ding, et al. Dalk: Dynamic co-augmentation of llms
and kg to answer alzheimer’s disease questions with scientific literature. arXiv preprint
_arXiv:2405.04819, 2024._
[225] Dawei Li, Shu Yang, Zhen Tan, Jae Young Baik, Sunkwon Yun, Joseph Lee, Aaron Chacko,
Bojian Hou, Duy Duong-Tran, Ying Ding, et al. Dalk: Dynamic co-augmentation of llms
and kg to answer alzheimer’s disease questions with scientific literature. arXiv preprint
_arXiv:2405.04819, 2024._
[226] Guanghua Li, Wensheng Lu, Wei Zhang, Defu Lian, Kezhong Lu, Rui Mao, Kai Shu, and
Hao Liao. Re-search for the truth: Multi-round retrieval-augmented large language models are
strong fake news detectors. arXiv preprint arXiv:2403.09747, 2024.
[227] Huayang Li, Yixuan Su, Deng Cai, Yan Wang, and Lemao Liu. A survey on retrievalaugmented text generation. arXiv preprint arXiv:2202.01110, 2022.
[228] Jing Li, Aixin Sun, Jianglei Han, and Chenliang Li. A survey on deep learning for named
entity recognition. IEEE transactions on knowledge and data engineering, 34(1):50–70, 2020.
[229] Liang Li, Ruiying Geng, Bowen Li, Can Ma, Yinliang Yue, Binhua Li, and Yongbin Li.
Graph-to-text generation with dynamic structure pruning. arXiv preprint arXiv:2209.07258,
2022.
[230] Linfeng Li, Peng Wang, Jun Yan, Yao Wang, Simin Li, Jinpeng Jiang, Zhe Sun, Buzhou
Tang, Tsung-Hui Chang, Shenghui Wang, et al. Real-world data medical knowledge graph:
construction and applications. Artificial intelligence in medicine, 103:101817, 2020.
[231] Miao Li, Jianzhong Qi, and Jey Han Lau. Compressed heterogeneous graph for abstractive multi-document summarization. In Proceedings of the AAAI Conference on Artificial
_Intelligence, volume 37, pages 13085–13093, 2023._
[232] Mufei Li, Siqi Miao, and Pan Li. Simple is effective: The roles of graphs and large language models in knowledge-graph-based retrieval-augmented generation. arXiv preprint
_arXiv:2410.20724, 2024._
[233] Ronghan Li, Lifang Wang, Zejun Jiang, Zhongtian Hu, Meng Zhao, and Xinyu Lu. Mutually improved dense retriever and gnn-based reader for arbitrary-hop open-domain question
answering. Neural Computing and Applications, 34(14):11831–11851, 2022.
[234] Shiyang Li, Yifan Gao, Haoming Jiang, Qingyu Yin, Zheng Li, Xifeng Yan, Chao Zhang,
and Bing Yin. Graph reasoning for question answering with triplet retrieval. arXiv preprint
_arXiv:2305.18742, 2023._
[235] Shuo Li et al. Traq: Trustworthy retrieval augmented question answering via conformal
[prediction. arXiv preprint arXiv:2307.04642, 2023. URL https://arxiv.org/abs/2307.](https://arxiv.org/abs/2307.04642)
```
04642.
```
[236] Tianle Li, Ge Zhang, Quy Duc Do, Xiang Yue, and Wenhu Chen. Long-context llms struggle
with long in-context learning. arXiv preprint arXiv:2404.02060, 2024.
[237] Wei Li, Xinyan Xiao, Jiachen Liu, Hua Wu, Haifeng Wang, and Junping Du. Leveraging graph
to improve abstractive multi-document summarization. arXiv preprint arXiv:2005.10043,
2020.
[238] Xingxuan Li, Ruochen Zhao, Yew Ken Chia, Bosheng Ding, Shafiq Joty, Soujanya Poria, and
Lidong Bing. Chain-of-knowledge: Grounding large language models via dynamic knowledge
adapting over heterogeneous sources. arXiv preprint arXiv:2305.13269, 2023.
-----
[239] Yongqi Li, Wenjie Li, and Liqiang Nie. A graph-guided multi-round retrieval method for
conversational open-domain question answering. arXiv preprint arXiv:2104.08443, 2021.
[240] Yongqi Li, Wenjie Li, and Liqiang Nie. Dynamic graph reasoning for conversational opendomain question answering. ACM Transactions on Information Systems (TOIS), 40(4):1–24,
2022.
[241] Yuan Li, Yixuan Zhang, and Lichao Sun. Metaagents: Simulating interactions of human
behaviors for llm-based task-oriented coordination via collaborative generative agents. arXiv
_preprint arXiv:2310.06500, 2023._
[242] Zekun Li, Zeyu Cui, Shu Wu, Xiaoyu Zhang, and Liang Wang. Fi-gnn: Modeling feature
interactions via graph neural networks for ctr prediction. In Proceedings of the 28th ACM
_international conference on information and knowledge management, pages 539–548, 2019._
[243] Zhuoqun Li, Xuanang Chen, Haiyang Yu, Hongyu Lin, Yaojie Lu, Qiaoyu Tang, Fei Huang,
Xianpei Han, Le Sun, and Yongbin Li. Structrag: Boosting knowledge intensive reasoning of
llms via inference-time hybrid information structurization. arXiv preprint arXiv:2410.08815,
2024.
[244] Zhuoyang Li, Liran Deng, Hui Liu, Qiaoqiao Liu, and Junzhao Du. Unioqa: A unified
framework for knowledge graph question answering with large language models. arXiv
_preprint arXiv:2406.02110, 2024._
[245] Jay Chiehen Liao and Cheng-Te Li. Tabgsl: Graph structure learning for tabular data prediction.
_arXiv preprint arXiv:2305.15843, 2023._
[246] Ruotong Liao, Xu Jia, Yangzhe Li, Yunpu Ma, and Volker Tresp. Gentkg: Generative
forecasting on temporal knowledge graph with large language models. In Findings of the
_Association for Computational Linguistics: NAACL 2024, pages 4303–4317, 2024._
[247] Wenlong Liao, Birgitte Bak-Jensen, Jayakrishnan Radhakrishna Pillai, Yuelong Wang, and
Yusen Wang. A review of graph neural networks and their applications in power systems.
_Journal of Modern Power Systems and Clean Energy, 10(2):345–360, 2021._
[248] Fangru Lin, Emanuele La Malfa, Valentin Hofmann, Elle Michelle Yang, Anthony Cohn,
and Janet B Pierrehumbert. Graph-enhanced large language models in asynchronous plan
reasoning. arXiv preprint arXiv:2402.02805, 2024.
[249] Bang Liu, Ting Zhang, Di Niu, Jinghong Lin, Kunfeng Lai, and Yu Xu. Matching long
text documents via graph convolutional networks. arXiv preprint arXiv:1802.07459, pages
2793–2799, 2018.
[250] Chang Liu, Xiaohui Xie, Xinggong Zhang, and Yong Cui. Large language models for
networking: Workflow, advances and challenges. arXiv preprint arXiv:2404.12901, 2024.
[251] Guangyi Liu, Yongqi Zhang, Yong Li, and Quanming Yao. Explore then determine: A gnn-llm
synergy framework for reasoning over knowledge graph. arXiv preprint arXiv:2406.01145,
2024.
[252] Haochen Liu, Song Wang, Yaochen Zhu, Yushun Dong, and Jundong Li. Knowledge graphenhanced large language models via path selection. arXiv preprint arXiv:2406.13862, 2024.
[[253] Jerry Liu. LlamaIndex, 11 2022. URL https://github.com/jerryjliu/llama_index.](https://github.com/jerryjliu/llama_index)
[254] Jiawei Liu, Cheng Yang, Zhiyuan Lu, Junze Chen, Yibo Li, Mengmei Zhang, Ting Bai, Yuan
Fang, Lichao Sun, Philip S Yu, et al. Towards graph foundation models: A survey and beyond.
_arXiv preprint arXiv:2310.11829, 2023._
[255] Jinzhe Liu, Xiangsheng Huang, Zhuo Chen, and Yin Fang. Drak: Unlocking molecular insights
with domain-specific retrieval-augmented knowledge in llms. Authorea Preprints, 2024.
[256] Nelson F Liu, Kevin Lin, John Hewitt, Ashwin Paranjape, Michele Bevilacqua, Fabio Petroni,
and Percy Liang. Lost in the middle: How language models use long contexts. Transactions
_of the Association for Computational Linguistics, 12:157–173, 2024._
-----
[257] Qi Liu, Maximilian Nickel, and Douwe Kiela. Hyperbolic graph neural networks. Advances
_in neural information processing systems, 32, 2019._
[258] Qijiong Liu, Nuo Chen, Tetsuya Sakai, and Xiao-Ming Wu. Once: Boosting content-based
recommendation with both open-and closed-source large language models. In Proceedings
_of the 17th ACM International Conference on Web Search and Data Mining, pages 452–461,_
2024.
[259] Tengfei Liu, Yongli Hu, Boyue Wang, Yanfeng Sun, Junbin Gao, and Baocai Yin. Hierarchical
graph convolutional networks for structured long document classification. IEEE transactions
_on neural networks and learning systems, 34(10):8071–8085, 2022._
[260] Wei Liu, Xiyan Fu, and Michael Strube. Modeling structural similarities between documents
for coherence assessment with graph convolutional networks. arXiv preprint arXiv:2306.06472,
2023.
[261] Weijie Liu, Peng Zhou, Zhe Zhao, Zhiruo Wang, Qi Ju, Haotang Deng, and Ping Wang.
K-bert: Enabling language representation with knowledge graph. In Proceedings of the AAAI
_Conference on Artificial Intelligence, volume 34, pages 2901–2908, 2020._
[262] Xiangyan Liu, Bo Lan, Zhiyuan Hu, Yang Liu, Zhicheng Zhang, Wenmeng Zhou, Fei Wang,
and Michael Shieh. Codexgraph: Bridging large language models and code repositories via
code graph databases. arXiv preprint arXiv:2408.03910, 2024.
[263] Yinhan Liu. Roberta: A robustly optimized bert pretraining approach. _arXiv preprint_
_arXiv:1907.11692, 2019._
[264] Yixin Liu, Ming Jin, Shirui Pan, Chuan Zhou, Yu Zheng, Feng Xia, and S Yu Philip. Graph
self-supervised learning: A survey. IEEE transactions on knowledge and data engineering, 35
(6):5879–5900, 2022.
[265] Yuyan Liu, Sirui Ding, Sheng Zhou, Wenqi Fan, and Qiaoyu Tan. Moleculargpt: Open
large language model (llm) for few-shot molecular property prediction. _arXiv preprint_
_arXiv:2406.12950, 2024._
[266] Xinwei Long, Jiali Zeng, Fandong Meng, Zhiyuan Ma, Kaiyan Zhang, Bowen Zhou, and Jie
Zhou. Generative multi-modal knowledge retrieval with large language models. In Proceedings
_of the AAAI Conference on Artificial Intelligence, volume 38, pages 18733–18741, 2024._
[267] Haohui Lu and Shahadat Uddin. A weighted patient network-based framework for predicting
chronic diseases using graph neural networks. Scientific reports, 11(1):22607, 2021.
[268] Qiaolin Lu, Jiayuan Ding, Lingxiao Li, Yi Chang, Jiliang Tang, and Xiaojie Qiu. Graph contrastive learning of subcellular-resolution spatial transcriptomics improves cell type annotation
and reveals critical molecular pathways. bioRxiv, pages 2024–03, 2024.
[269] Zhiyong Lu. Pubmed and beyond: a survey of web tools for searching biomedical literature.
_Database, 2011:baq036, 2011._
[270] Haoran Luo, Zichen Tang, Shiyao Peng, Yikai Guo, Wentai Zhang, Chenghao Ma, Guanting
Dong, Meina Song, Wei Lin, et al. Chatkbqa: A generate-then-retrieve framework for
knowledge base question answering with fine-tuned large language models. arXiv preprint
_arXiv:2310.08975, 2023._
[271] LINHAO LUO, Yuan-Fang Li, Reza Haf, and Shirui Pan. Reasoning on graphs: Faithful and
interpretable large language model reasoning. In The Twelfth International Conference on
_Learning Representations._
[272] Linhao Luo, Yuan-Fang Li, Gholamreza Haffari, and Shirui Pan. Reasoning on graphs: Faithful
and interpretable large language model reasoning. arXiv preprint arXiv:2310.01061, 2023.
[273] Linhao Luo, Zicheng Zhao, Chen Gong, Gholamreza Haffari, and Shirui Pan. Graphconstrained reasoning: Faithful reasoning on knowledge graphs with large language models.
_arXiv preprint arXiv:2410.13080, 2024._
-----
[274] Shitong Luo, Chence Shi, Minkai Xu, and Jian Tang. Predicting molecular conformation
via dynamic graph score matching. Advances in Neural Information Processing Systems, 34:
19784–19795, 2021.
[275] Zihan Luo, Xiran Song, Hong Huang, Jianxun Lian, Chenhao Zhang, Jinqi Jiang, Xing Xie,
and Hai Jin. Graphinstruct: Empowering large language models with graph understanding and
reasoning capability. arXiv preprint arXiv:2403.04483, 2024.
[276] Hanjia Lyu, Song Jiang, Hanqing Zeng, Yinglong Xia, Qifan Wang, Si Zhang, Ren Chen,
Christopher Leung, Jiajie Tang, and Jiebo Luo. Llm-rec: Personalized recommendation via
prompting large language models. arXiv preprint arXiv:2307.15780, 2023.
[277] Qiyao Ma, Menglin Yang, Mingxuan Ju, Tong Zhao, Neil Shah, and Rex Ying. Harec:
Hyperbolic graph-llm alignment for exploration and exploitation in recommender systems.
_arXiv preprint arXiv:2411.13865, 2024._
[278] Shengjie Ma, Chengjin Xu, Xuhui Jiang, Muzhi Li, Huaren Qu, and Jian Guo. Think-on-graph
2.0: Deep and interpretable large language model reasoning with knowledge graph-guided
retrieval. arXiv preprint arXiv:2407.10805, 2024.
[279] Yao Ma and Jiliang Tang. Deep learning on graphs. Cambridge University Press, 2021.
[280] Yao Ma, Xiaorui Liu, Neil Shah, and Jiliang Tang. Is homophily a necessity for graph neural
networks? arXiv preprint arXiv:2106.06134, 2021.
[281] Andrzej Ma´ckiewicz and Waldemar Ratajczak. Principal components analysis (pca). Comput_ers & Geosciences, 19(3):303–342, 1993._
[282] Aman Madaan, Dheeraj Rajagopal, Niket Tandon, Yiming Yang, and Eduard Hovy. Could
you give me a hint? generating inference graphs for defeasible reasoning. arXiv preprint
_arXiv:2105.05418, 2021._
[283] Ali Madani, Ben Krause, Eric R Greene, Subu Subramanian, Benjamin P Mohr, James M
Holton, Jose Luis Olmos, Caiming Xiong, Zachary Z Sun, Richard Socher, et al. Large
language models generate functional protein sequences across diverse families. Nature Biotech_nology, 41(8):1099–1106, 2023._
[284] Himanshu Maheshwari, Sambaran Bandyopadhyay, Aparna Garimella, and Anandhavelu
Natarajan. Presentations are not always linear! gnn meets llm for document-to-presentation
transformation with attribution. arXiv preprint arXiv:2405.13095, 2024.
[285] Haitao Mao, Zhikai Chen, Wei Jin, Haoyu Han, Yao Ma, Tong Zhao, Neil Shah, and Jiliang
Tang. Demystifying structural disparity in graph neural networks: Can one size fit all?
_Advances in neural information processing systems, 36, 2024._
[286] Haitao Mao, Zhikai Chen, Wenzhuo Tang, Jianan Zhao, Yao Ma, Tong Zhao, Neil Shah,
Michael Galkin, and Jiliang Tang. Graph foundation models. arXiv preprint arXiv:2402.02216,
2024.
[287] Xuting Mao, Hao Sun, Xiaoqian Zhu, and Jianping Li. Financial fraud detection using the
related-party transaction knowledge graph. Procedia Computer Science, 199:733–740, 2022.
[288] Vaibhav Mavi, Anubhav Jangra, Jatowt Adam, et al. Multi-hop question answering. Founda_tions and Trends® in Information Retrieval, 17(5):457–586, 2024._
[289] Costas Mavromatis and George Karypis. Gnn-rag: Graph neural retrieval for large language
model reasoning. arXiv preprint arXiv:2405.20139, 2024.
[290] Julian McAuley, Christopher Targett, Qinfeng Shi, and Anton Van Den Hengel. Image-based
recommendations on styles and substitutes. In Proceedings of the 38th international ACM
_SIGIR conference on research and development in information retrieval, pages 43–52, 2015._
[291] Miller McPherson, Lynn Smith-Lovin, and James M Cook. Birds of a feather: Homophily in
social networks. Annual review of sociology, 27(1):415–444, 2001.
-----
[292] Xin Mei, Xiaoyan Cai, Libin Yang, and Nanxin Wang. Graph transformer networks based text
representation. Neurocomputing, 463:91–100, 2021.
[293] Yu Meng, Yunyi Zhang, Jiaxin Huang, Xuan Wang, Yu Zhang, Heng Ji, and Jiawei Han.
Distantly-supervised named entity recognition with noise-robust learning and language model
augmented self-training. arXiv preprint arXiv:2109.05003, 2021.
[294] R Mihalcea. Graph-based natural language processing and information retrieval. Cambridge
University Press, 2011.
[295] Eleni P Mimitou, Anthony Cheng, Antonino Montalbano, Stephanie Hao, Marlon Stoeckius,
Mateusz Legut, Timothy Roush, Alberto Herrera, Efthymia Papalexi, Zhengqing Ouyang, et al.
Multiplexed detection of proteins, transcriptomes, clonotypes and crispr perturbations in single
cells. Nature methods, 16(5):409–412, 2019.
[296] Erxue Min, Runfa Chen, Yatao Bian, Tingyang Xu, Kangfei Zhao, Wenbing Huang, Peilin
Zhao, Junzhou Huang, Sophia Ananiadou, and Yu Rong. Transformer for graphs: An overview
from architecture perspective. arXiv preprint arXiv:2202.08455, 2022.
[297] Sewon Min, Danqi Chen, Luke Zettlemoyer, and Hannaneh Hajishirzi. Knowledge guided text
retrieval and reading for open domain question answering. arXiv preprint arXiv:1911.03868,
2019.
[298] Marco Minici, Federico Cinus, Corrado Monti, Francesco Bonchi, and Giuseppe Manco.
Cascade-based echo chamber detection. In Proceedings of the 31st ACM International
_Conference on Information & Knowledge Management, pages 1511–1520, 2022._
[299] Dylan Molho, Jiayuan Ding, Wenzhuo Tang, Zhaoheng Li, Hongzhi Wen, Yixin Wang, Julian
Venegas, Wei Jin, Renming Liu, Runze Su, et al. Deep learning in single-cell analysis. ACM
_Transactions on Intelligent Systems and Technology, 15(3):1–62, 2024._
[300] Sai Munikoti, Anurag Acharya, Sridevi Wagle, and Sameera Horawalavithana. Atlantic:
Structure-aware retrieval-augmented language model for interdisciplinary science. arXiv
_preprint arXiv:2311.12289, 2023._
[301] Lino Murali, G Gopakumar, Daleesha M Viswanathan, and Prema Nedungadi. Towards
electronic health record-based medical knowledge graph construction, completion, and applications: A literature study. Journal of biomedical informatics, 143:104403, 2023.
[302] Guoshun Nan, Zhijiang Guo, Ivan Sekuli´c, and Wei Lu. Reasoning with latent structure
refinement for document-level relation extraction. arXiv preprint arXiv:2005.06312, 2020.
[303] Zara Nasar, Syed Waqar Jaffry, and Muhammad Kamran Malik. Named entity recognition and
relation extraction: State-of-the-art. ACM Computing Surveys (CSUR), 54(1):1–39, 2021.
[304] Mark Newman. Networks. Oxford university press, 2018.
[305] Bo Ni, Yu Wang, Lu Cheng, Erik Blasch, and Tyler Derr. Towards trustworthy knowledge
graph reasoning: An uncertainty aware perspective. arXiv preprint arXiv:2410.08985, 2024.
[306] Jianmo Ni, Jiacheng Li, and Julian McAuley. Justifying recommendations using distantlylabeled reviews and fine-grained aspects. In Proceedings of the 2019 conference on empirical
_methods in natural language processing and the 9th international joint conference on natural_
_language processing (EMNLP-IJCNLP), pages 188–197, 2019._
[307] Yuxiang Nie, Heyan Huang, Wei Wei, and Xian-Ling Mao. Capturing global structural
information in long document question answering with compressive graph selector network.
_arXiv preprint arXiv:2210.05499, 2022._
[308] Mengjia Niu, Hao Li, Jie Shi, Hamed Haddadi, and Fan Mo. Mitigating hallucinations in
large language models via self-refinement-enhanced knowledge retrieval. arXiv preprint
_arXiv:2405.06545, 2024._
-----
[309] Barlas Oguz, Xilun Chen, Vladimir Karpukhin, Stan Peshterliev, Dmytro Okhonko, Michael
Schlichtkrull, Sonal Gupta, Yashar Mehdad, and Scott Yih. Unik-qa: Unified representations
of structured and unstructured knowledge for open-domain question answering. arXiv preprint
_arXiv:2012.14610, 2020._
[310] Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll Wainwright, Pamela Mishkin,
Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. Training language models
to follow instructions with human feedback. Advances in neural information processing
_systems, 35:27730–27744, 2022._
[311] Damian Owerko, Fernando Gama, and Alejandro Ribeiro. Optimal power flow using graph
neural networks. In ICASSP 2020-2020 IEEE International Conference on Acoustics, Speech
_and Signal Processing (ICASSP), pages 5930–5934. IEEE, 2020._
[312] Vardaan Pahuja, Boshi Wang, Hugo Latapie, Jayanth Srinivasa, and Yu Su. A retrieve-and-read
framework for knowledge graph link prediction. In Proceedings of the 32nd ACM International
_Conference on Information and Knowledge Management, pages 1992–2002, 2023._
[313] Liangming Pan, Yuxi Xie, Yansong Feng, Tat-Seng Chua, and Min-Yen Kan. Semantic graphs
for generating deep questions. arXiv preprint arXiv:2004.12704, 2020.
[314] Sarah Parisot, Sofia Ira Ktena, Enzo Ferrante, Matthew Lee, Ricardo Guerrero, Ben Glocker,
and Daniel Rueckert. Disease prediction using graph convolutional networks: application to
autism spectrum disorder and alzheimer’s disease. Medical image analysis, 48:117–130, 2018.
[315] Dae Hoon Park, Hyun Duk Kim, ChengXiang Zhai, and Lifan Guo. Retrieval of relevant
opinion sentences for new products. In Proceedings of the 38th International ACM SIGIR
_Conference on Research and Development in Information Retrieval, pages 393–402, 2015._
[316] Jinyoung Park, Ameen Patel, Omar Zia Khan, Hyunwoo J Kim, and Joo-Kyung Kim. Graphguided reasoning for multi-hop question answering in large language models. arXiv preprint
_arXiv:2311.09762, 2023._
[317] Sachin Pawar, Girish K Palshikar, and Pushpak Bhattacharyya. Relation extraction: A survey.
_arXiv preprint arXiv:1712.05191, 2017._
[318] Alexander R Pelletier, Joseph Ramirez, Irsyad Adam, Simha Sankar, Yu Yan, Ding Wang,
Dylan Steinecke, Wei Wang, and Peipei Ping. Explainable biomedical hypothesis generation via retrieval augmented generation enabled large language models. arXiv preprint
_arXiv:2407.12888, 2024._
[319] Boci Peng, Yun Zhu, Yongchao Liu, Xiaohe Bo, Haizhou Shi, Chuntao Hong, Yan Zhang,
and Siliang Tang. Graph retrieval-augmented generation: A survey. _arXiv preprint_
_arXiv:2408.08921, 2024._
[320] Zhuoyi Peng and Yi Yang. Connecting the dots: Inferring patent phrase similarity with
retrieved phrase graphs. arXiv preprint arXiv:2403.16265, 2024.
[321] Bryan Perozzi, Rami Al-Rfou, and Steven Skiena. Deepwalk: Online learning of social
representations. In Proceedings of the 20th ACM SIGKDD international conference on
_Knowledge discovery and data mining, pages 701–710, 2014._
[322] Joseph Priestley. “The” History and Present State of Electricity: With Original Experiments.
C. Bathurst and T. Lowndes, in Fleet-Street, J. Rivington and J. Johnson, in ..., 1775.
[323] Chen Qian, Xiaochang Li, Qineng Wang, Gang Zhou, and Huajie Shao. Netbench: A largescale and comprehensive network traffic benchmark dataset for foundation models. arXiv
_preprint arXiv:2403.10319, 2024._
[324] Jianwei Qian, Xiang-Yang Li, Chunhong Zhang, Linlin Chen, Taeho Jung, and Junze Han.
Social network de-anonymization and privacy inference with knowledge graph model. IEEE
_Transactions on Dependable and Secure Computing, 16(4):679–692, 2017._
-----
[325] Tao Qin and Tie-Yan Liu. Introducing letor 4.0 datasets. arXiv preprint arXiv:1306.2597,
2013.
[326] Lin Qiu, Yunxuan Xiao, Yanru Qu, Hao Zhou, Lei Li, Weinan Zhang, and Yong Yu. Dynamically fused graph network for multi-hop reasoning. In Proceedings of the 57th annual meeting
_of the association for computational linguistics, pages 6140–6150, 2019._
[327] Ruihong Qiu, Zi Huang, Jingjing Li, and Hongzhi Yin. Exploiting cross-session information for
session-based recommendation with graph neural networks. ACM Transactions on Information
_Systems (TOIS), 38(3):1–23, 2020._
[328] Victor Quach, Adam Fisch, Tal Schuster, Adam Yala, Jae Ho Sohn, Tommi S. Jaakkola, and
[Regina Barzilay. Conformal language modeling, 2024. URL https://arxiv.org/abs/](https://arxiv.org/abs/2306.10193)
```
2306.10193.
```
[329] Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena,
Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified
text-to-text transformer. Journal of machine learning research, 21(140):1–67, 2020.
[330] Moqsadur Rahman, Krish O Piryani, Aaron M Sanchez, Sai Munikoti, Luis De La Torre,
Maxwell S Levin, Monika Akbar, Mahmud Hossain, Monowar Hasan, and Mahantesh Halappanavar. Retrieval augmented generation for robust cyber defense. Technical report, Pacific
Northwest National Laboratory (PNNL), Richland, WA (United States), 2024.
[331] Ashwin Ram, Yigit Ege Bayiz, Arash Amini, Mustafa Munir, and Radu Marculescu. Credirag:
Network-augmented credibility-based retrieval for misinformation detection in reddit. arXiv
_preprint arXiv:2410.12061, 2024._
[332] Gowtham Ramesh, Makesh Sreedhar, and Junjie Hu. Single sequence prediction over reasoning
graphs for multi-hop qa. arXiv preprint arXiv:2307.00335, 2023.
[333] Juan Ramos et al. Using tf-idf to determine word relevance in document queries. In Proceedings
_of the first instructional conference on machine learning, volume 242, pages 29–48. Citeseer,_
2003.
[334] Jiahua Rao, Xiang Zhou, Yutong Lu, Huiying Zhao, and Yuedong Yang. Imputing single-cell
rna-seq data by combining graph convolution and autoencoder neural networks. Iscience, 24
(5), 2021.
[335] Susie Xi Rao, Shuai Zhang, Zhichao Han, Zitao Zhang, Wei Min, Zhiyao Chen, Yinan Shan,
Yang Zhao, and Ce Zhang. xfraud: explainable fraud transaction detection. arXiv preprint
_arXiv:2011.12193, 2020._
[336] Leonardo FR Ribeiro, Martin Schmitt, Hinrich Schütze, and Iryna Gurevych. Investigating
pretrained language models for graph-to-text generation. arXiv preprint arXiv:2007.08426,
2020.
[337] Stephen Robertson, Hugo Zaragoza, and Michael Taylor. Simple bm25 extension to multiple
weighted fields. In Proceedings of the thirteenth ACM international conference on Information
_and knowledge management, pages 42–49, 2004._
[338] Stephen Robertson, Hugo Zaragoza, et al. The probabilistic relevance framework: Bm25 and
beyond. Foundations and Trends® in Information Retrieval, 3(4):333–389, 2009.
[339] Joshua Robinson, Rishabh Ranjan, Weihua Hu, Kexin Huang, Jiaqi Han, Alejandro Dobles,
Matthias Fey, Jan E Lenssen, Yiwen Yuan, Zecheng Zhang, et al. Relbench: A benchmark for
deep learning on relational databases. arXiv preprint arXiv:2407.20060, 2024.
[340] Ryan Rossi and Nesreen Ahmed. The network data repository with interactive graph analytics
and visualization. In Proceedings of the AAAI conference on artificial intelligence, volume 29,
2015.
[341] Krzysztof Rusek, José Suárez-Varela, Paul Almasan, Pere Barlet-Ros, and Albert CabellosAparicio. Routenet: Leveraging graph neural networks for network modeling and optimization
in sdn. IEEE Journal on Selected Areas in Communications, 38(10):2260–2270, 2020.
-----
[342] Swarnadeep Saha, Prateek Yadav, Lisa Bauer, and Mohit Bansal. Explagraphs: An explanation
graph generation task for structured commonsense reasoning. arXiv preprint arXiv:2104.07644,
2021.
[343] Maria Sahakyan, Zeyar Aung, and Talal Rahwan. Explainable artificial intelligence for tabular
data: A survey. IEEE access, 9:135392–135422, 2021.
[344] Sunil Kumar Sahu, Fenia Christopoulou, Makoto Miwa, and Sophia Ananiadou. Inter-sentence
relation extraction with document-level graph convolutional neural network. arXiv preprint
_arXiv:1906.04684, 2019._
[345] Md Sadman Sakib and Yu Sun. Consolidating trees of robotic plans generated using large
language models to improve reliability. arXiv preprint arXiv:2401.07868, 2024.
[346] Alireza Salemi, Sheshera Mysore, Michael Bendersky, and Hamed Zamani. Lamp: When
large language models meet personalization. arXiv preprint arXiv:2304.11406, 2023.
[347] Diego Sanmartin. Kg-rag: Bridging the gap between knowledge and creativity. arXiv preprint
_arXiv:2405.12035, 2024._
[348] Abulhair Saparov and He He. Language models are greedy reasoners: A systematic formal
analysis of chain-of-thought. arXiv preprint arXiv:2210.01240, 2022.
[349] Vıctor Garcia Satorras, Emiel Hoogeboom, and Max Welling. E (n) equivariant graph neural
networks. In International conference on machine learning, pages 9323–9332. PMLR, 2021.
[350] Michael Schlichtkrull, Thomas N Kipf, Peter Bloem, Rianne Van Den Berg, Ivan Titov, and
Max Welling. Modeling relational data with graph convolutional networks. In The semantic
_web: 15th international conference, ESWC 2018, Heraklion, Crete, Greece, June 3–7, 2018,_
_proceedings 15, pages 593–607. Springer, 2018._
[351] Michael Schuhmacher and Simone Paolo Ponzetto. Knowledge-based graph document modeling. In Proceedings of the 7th ACM international conference on Web search and data mining,
pages 543–552, 2014.
[352] Xin Shao, Haihong Yang, Xiang Zhuang, Jie Liao, Penghui Yang, Junyun Cheng, Xiaoyan
Lu, Huajun Chen, and Xiaohui Fan. scdeepsort: a pre-trained cell-type annotation method for
single-cell transcriptomics using deep learning with a weighted graph neural network. Nucleic
_acids research, 49(21):e122–e122, 2021._
[353] Vasu Sharma, Harsh Vardhan Sharma, Ankita Bishnu, and Labhesh Patel. Cyclegen: Cyclic
consistency based product review generator from attributes. In Proceedings of the 11th
_International Conference on Natural Language Generation, pages 426–430, 2018._
[354] Ahsan Shehzad, Feng Xia, Shagufta Abid, Ciyuan Peng, Shuo Yu, Dongyu Zhang, and Karin
Verspoor. Graph transformers: A survey. arXiv preprint arXiv:2407.09777, 2024.
[355] Yongliang Shen, Kaitao Song, Xu Tan, Wenqi Zhang, Kan Ren, Siyu Yuan, Weiming Lu,
Dongsheng Li, and Yueting Zhuang. Taskbench: Benchmarking large language models for
task automation. arXiv preprint arXiv:2311.18760, 2023.
[356] Yongliang Shen, Kaitao Song, Xu Tan, Dongsheng Li, Weiming Lu, and Yueting Zhuang.
Hugginggpt: Solving ai tasks with chatgpt and its friends in hugging face. Advances in Neural
_Information Processing Systems, 36, 2024._
[357] Nino Shervashidze, Pascal Schweitzer, Erik Jan Van Leeuwen, Kurt Mehlhorn, and Karsten M
Borgwardt. Weisfeiler-lehman graph kernels. Journal of Machine Learning Research, 12(9),
2011.
[358] Freda Shi, Xinyun Chen, Kanishka Misra, Nathan Scales, David Dohan, Ed H Chi, Nathanael
Schärli, and Denny Zhou. Large language models can be easily distracted by irrelevant context.
In International Conference on Machine Learning, pages 31210–31227. PMLR, 2023.
[359] Peng Shi and Jimmy Lin. Simple bert models for relation extraction and semantic role labeling.
_arXiv preprint arXiv:1904.05255, 2019._
-----
[360] Yucheng Shi, Qiaoyu Tan, Xuansheng Wu, Shaochen Zhong, Kaixiong Zhou, and Ninghao
Liu. Retrieval-enhanced knowledge editing for multi-hop question answering in language
models. arXiv preprint arXiv:2403.19631, 2024.
[361] Harry Shomer, Yao Ma, Haitao Mao, Juanhui Li, Bo Wu, and Jiliang Tang. LPFormer: An
adaptive graph transformer for link prediction. In Proceedings of the 30th ACM SIGKDD
_Conference on Knowledge Discovery and Data Mining, pages 2686–2698, 2024._
[362] Robik Shrestha, Yang Zou, Qiuyu Chen, Zhiheng Li, Yusheng Xie, and Siqi Deng. Fairrag:
Fair human generation via fair retrieval augmentation. In Proceedings of the IEEE/CVF
_Conference on Computer Vision and Pattern Recognition, pages 11996–12005, 2024._
[363] Ravid Shwartz-Ziv and Amitai Armon. Tabular data: Deep learning is not all you need.
_Information Fusion, 81:84–90, 2022._
[364] Karandeep Singh, Yu-Che Tsai, Cheng-Te Li, Meeyoung Cha, and Shou-De Lin. Graphfc:
Customs fraud detection with label scarcity. In Proceedings of the 32nd ACM International
_Conference on Information and Knowledge Management, pages 4829–4835, 2023._
[365] Ananya Singha, José Cambronero, Sumit Gulwani, Vu Le, and Chris Parnin. Tabular representation, noisy operators, and impacts on table structure understanding tasks in llms. arXiv
_preprint arXiv:2310.10358, 2023._
[366] Karthik Soman, Peter W Rose, John H Morris, Rabia E Akbas, Brett Smith, Braian Peetoom,
Catalina Villouta-Reyes, Gabriel Cerono, Yongmei Shi, Angela Rizk-Jackson, et al. Biomedical
knowledge graph-optimized prompt generation for large language models. Bioinformatics, 40
(9):btae560, 2024.
[367] Sheetal S Sonawane and Parag A Kulkarni. Graph based representation and analysis of text
document: A survey of techniques. International Journal of Computer Applications, 96(19),
2014.
[368] Chan Hee Song, Jiaman Wu, Clayton Washington, Brian M Sadler, Wei-Lun Chao, and Yu Su.
Llm-planner: Few-shot grounded planning for embodied agents with large language models. In
_Proceedings of the IEEE/CVF International Conference on Computer Vision, pages 2998–3009,_
2023.
[369] Linfeng Song, Zhiguo Wang, Mo Yu, Yue Zhang, Radu Florian, and Daniel Gildea. Exploring
graph-structured passage representation for multi-hop reading comprehension with graph
neural networks. arXiv preprint arXiv:1809.02040, 2018.
[370] Linfeng Song, Yue Zhang, Zhiguo Wang, and Daniel Gildea. A graph-to-sequence model for
amr-to-text generation. arXiv preprint arXiv:1805.02473, 2018.
[371] Qianqian Song and Jing Su. Dstg: deconvoluting spatial transcriptomics data through graphbased artificial intelligence. Briefings in bioinformatics, 22(5):bbaa414, 2021.
[372] Yifan Song, Weimin Xiong, Dawei Zhu, Wenhao Wu, Han Qian, Mingbo Song, Hailiang
Huang, Cheng Li, Ke Wang, Rong Yao, et al. Restgpt: Connecting large language models with
real-world restful apis. arXiv preprint arXiv:2306.06624, 2023.
[373] Robyn Speer, Joshua Chin, and Catherine Havasi. Conceptnet 5.5: An open multilingual
graph of general knowledge. In Proceedings of the AAAI conference on artificial intelligence,
volume 31, 2017.
[374] Indro Spinelli, Simone Scardapane, and Aurelio Uncini. Missing data imputation with
adversarially-trained graph convolutional networks. Neural Networks, 129:249–260, 2020.
[375] Tim Stuart and Rahul Satija. Integrative single-cell analysis. Nature reviews genetics, 20(5):
257–272, 2019.
[376] Jiayuan Su, Jing Luo, Hongwei Wang, and Lu Cheng. Api is enough: Conformal prediction for
[large language models without logit-access, 2024. URL https://arxiv.org/abs/2403.](https://arxiv.org/abs/2403.01216)
```
01216.
```
-----
[377] Ying Su, Jipeng Zhang, Yangqiu Song, and Tong Zhang. Pipenet: Question answering with
semantic pruning over knowledge graphs. arXiv preprint arXiv:2401.17536, 2024.
[378] Yuan Sui, Mengyu Zhou, Mingjie Zhou, Shi Han, and Dongmei Zhang. Table meets llm: Can
large language models understand structured table data? a benchmark and empirical study.
In Proceedings of the 17th ACM International Conference on Web Search and Data Mining,
pages 645–654, 2024.
[379] Haitian Sun, Bhuwan Dhingra, Manzil Zaheer, Kathryn Mazaitis, Ruslan Salakhutdinov, and
William Cohen. Open domain question answering using early fusion of knowledge bases
and text. In Proceedings of the 2018 Conference on Empirical Methods in Natural Language
_Processing, pages 4231–4242, 2018._
[380] Haitian Sun, Tania Bedrax-Weiss, and William Cohen. Pullnet: Open domain question
answering with iterative retrieval on knowledge bases and text. In Proceedings of the 2019
_Conference on Empirical Methods in Natural Language Processing and the 9th International_
_Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 2380–2390,_
2019.
[381] Jiashuo Sun, Chengjin Xu, Lumingyuan Tang, Saizhuo Wang, Chen Lin, Yeyun Gong, Lionel
Ni, Heung-Yeung Shum, and Jian Guo. Think-on-graph: Deep and responsible reasoning
of large language model on knowledge graph. In The Twelfth International Conference on
_Learning Representations._
[382] Yueqing Sun, Qi Shi, Le Qi, and Yu Zhang. Jointlk: Joint reasoning with language models and
knowledge graphs for commonsense question answering. arXiv preprint arXiv:2112.02732,
2021.
[383] Zhiqing Sun, Zhi-Hong Deng, Jian-Yun Nie, and Jian Tang. Rotate: Knowledge graph
embedding by relational rotation in complex space. In International Conference on Learning
_Representations, 2019._
[384] John Sweller. Cognitive load during problem solving: Effects on learning. Cognitive science,
12(2):257–285, 1988.
[385] Damian Szklarczyk, Annika L Gable, David Lyon, Alexander Junge, Stefan Wyder, Jaime
Huerta-Cepas, Milan Simonovic, Nadezhda T Doncheva, John H Morris, Peer Bork, Lars J
Jensen, and Christian von Mering. STRING v11: Protein–protein association networks with
increased coverage, supporting functional discovery in genome-wide experimental datasets.
_Nucleic Acids Research, 47(D1):D607–D613, 2019. ISSN 0305-1048, 1362-4962. doi:_
10.1093/nar/gky1131.
[386] Hengliang Tang, Yuan Mi, Fei Xue, and Yang Cao. An integration model based on graph
convolutional network for text classification. IEEE Access, 8:148865–148876, 2020.
[387] Wenzhuo Tang, Hongzhi Wen, Renming Liu, Jiayuan Ding, Wei Jin, Yuying Xie, Hui Liu,
and Jiliang Tang. Single-cell multimodal prediction via transformers. In Proceedings of the
_32nd ACM International Conference on Information and Knowledge Management, pages_
2422–2431, 2023.
[388] Wenzhuo Tang, Haitao Mao, Danial Dervovic, Ivan Brugere, Saumitra Mishra, Yuying Xie,
and Jiliang Tang. Cross-domain graph data scaling: A showcase with diffusion models. arXiv
_preprint arXiv:2406.01899, 2024._
[389] Dhaval Taunk, Lakshya Khanna, Siri Venkata Pavan Kumar Kandru, Vasudeva Varma, Charu
Sharma, and Makarand Tapaswi. Grapeqa: Graph augmentation and pruning to enhance
question-answering. In Companion Proceedings of the ACM Web Conference 2023, pages
1138–1144, 2023.
[390] Gemini Team, Rohan Anil, Sebastian Borgeaud, Yonghui Wu, Jean-Baptiste Alayrac, Jiahui
Yu, Radu Soricut, Johan Schalkwyk, Andrew M Dai, Anja Hauth, et al. Gemini: a family of
highly capable multimodal models. arXiv preprint arXiv:2312.11805, 2023.
-----
[391] Gemma Team, Thomas Mesnard, Cassidy Hardin, Robert Dadashi, Surya Bhupatiraju, Shreya
Pathak, Laurent Sifre, Morgane Rivière, Mihir Sanjay Kale, Juliette Love, et al. Gemma: Open
models based on gemini research and technology. arXiv preprint arXiv:2403.08295, 2024.
[392] Mokanarangan Thayaparan, Marco Valentino, Viktor Schlegel, and André Freitas. Identifying
supporting facts for multi-hop question answering with document graph networks. arXiv
_preprint arXiv:1910.00290, 2019._
[393] Arun James Thirunavukarasu, Darren Shu Jeng Ting, Kabilan Elangovan, Laura Gutierrez,
Ting Fang Tan, and Daniel Shu Wei Ting. Large language models in medicine. Nature
_medicine, 29(8):1930–1940, 2023._
[394] Maung Thway, Jose Recatala-Gomez, Fun Siong Lim, Kedar Hippalgaonkar, and Leonard
W. T. Ng. Battling botpoop using genai for higher education: A study of a retrieval augmented
generation chatbots impact on learning. arXiv preprint arXiv:2312.10997, 2023.
[395] Yijun Tian, Huan Song, Zichen Wang, Haozhu Wang, Ziqing Hu, Fang Wang, Nitesh V
Chawla, and Panpan Xu. Graph neural prompting with large language models. In Proceedings
_of the AAAI Conference on Artificial Intelligence, volume 38, pages 19080–19088, 2024._
[396] Yu Tian, Yuhao Yang, Xudong Ren, Pengfei Wang, Fangzhao Wu, Qian Wang, and Chenliang
Li. Joint knowledge pruning and recurrent graph convolution for news recommendation. In
_Proceedings of the 44th international ACM SIGIR conference on research and development in_
_information retrieval, pages 51–60, 2021._
[397] SM Tonmoy, SM Zaman, Vinija Jain, Anku Rani, Vipula Rawte, Aman Chadha, and Amitava
Das. A comprehensive survey of hallucination mitigation techniques in large language models.
_arXiv preprint arXiv:2401.01313, 2024._
[398] Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei,
Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, et al. Llama 2: Open
foundation and fine-tuned chat models. arXiv preprint arXiv:2307.09288, 2023.
[399] Harsh Trivedi, Niranjan Balasubramanian, Tushar Khot, and Ashish Sabharwal. Interleaving
retrieval with chain-of-thought reasoning for knowledge-intensive multi-step questions. arXiv
_preprint arXiv:2212.10509, 2022._
[400] George Tsatsaronis, Georgios Balikas, Prodromos Malakasiotis, Ioannis Partalas, Matthias
Zschunke, Michael R Alvers, Dirk Weissenborn, Anastasia Krithara, Sergios Petridis, Dimitris
Polychronopoulos, et al. An overview of the bioasq large-scale biomedical semantic indexing
and question answering competition. BMC bioinformatics, 16:1–28, 2015.
[401] Ming Tu, Kevin Huang, Guangtao Wang, Jing Huang, Xiaodong He, and Bowen Zhou. Select,
answer and explain: Interpretable multi-hop reading comprehension over multiple documents.
In Proceedings of the AAAI conference on artificial intelligence, volume 34, pages 9073–9080,
2020.
[402] Petar Veliˇckovi´c, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro Lio, and
Yoshua Bengio. Graph attention networks. arXiv preprint arXiv:1710.10903, 2017.
[403] S Vichy N Vishwanathan, Nicol N Schraudolph, Risi Kondor, and Karsten M Borgwardt.
Graph kernels. The Journal of Machine Learning Research, 11:1201–1242, 2010.
[404] Denny Vrandeˇci´c and Markus Krötzsch. Wikidata: a free collaborative knowledgebase.
_Communications of the ACM, 57(10):78–85, 2014._
[405] Sridevi Wagle, Sai Munikoti, Anurag Acharya, Sara Smith, and Sameera Horawalavithana.
Empirical evaluation of uncertainty quantification in retrieval-augmented language models for
science. arXiv preprint arXiv:2311.09358, 2023.
[406] Chaojie Wang, Yishi Xu, Zhong Peng, Chenxi Zhang, Bo Chen, Xinrun Wang, Lei Feng, and
Bo An. keqing: knowledge-based question answering is a nature chain-of-thought mentor of
llm. arXiv preprint arXiv:2401.00426, 2023.
-----
[407] Chen Wang, Dengji Zhou, Xiaoguo Wang, Song Liu, Tiemin Shao, Chongyuan Shui, and
Jun Yan. Multiscale graph based spatio-temporal graph convolutional network for energy
consumption prediction of natural gas transmission process. Energy, 307:132489, 2024.
[408] Danqing Wang, Pengfei Liu, Yining Zheng, Xipeng Qiu, and Xuanjing Huang. Heterogeneous graph neural networks for extractive document summarization. _arXiv preprint_
_arXiv:2004.12393, 2020._
[409] Difeng Wang, Wei Hu, Ermei Cao, and Weijian Sun. Global-to-local neural networks for
document-level relation extraction. arXiv preprint arXiv:2009.10359, 2020.
[410] Dongsheng Wang, Zhiqiang Ma, Armineh Nourbakhsh, Kang Gu, and Sameena Shah. Docgraphlm: Documental graph language model for information extraction. In Proceedings of
_the 46th International ACM SIGIR Conference on Research and Development in Information_
_Retrieval, pages 1944–1948, 2023._
[411] Fei Wang, Kexuan Sun, Muhao Chen, Jay Pujara, and Pedro Szekely. Retrieving complex tables
with multi-granular graph representation learning. In Proceedings of the 44th International
_ACM SIGIR Conference on Research and Development in Information Retrieval, pages 1472–_
1482, 2021.
[412] Fei Wang, Xingchen Wan, Ruoxi Sun, Jiefeng Chen, and Sercan Ö Arık. Astute rag: Overcoming imperfect retrieval augmentation and knowledge conflicts for large language models.
_arXiv preprint arXiv:2410.07176, 2024._
[413] Heng Wang, Shangbin Feng, Tianxing He, Zhaoxuan Tan, Xiaochuang Han, and Yulia
Tsvetkov. Can language models solve graph problems in natural language? _Advances_
_in Neural Information Processing Systems, 36, 2024._
[414] Juexin Wang, Anjun Ma, Yuzhou Chang, Jianting Gong, Yuexu Jiang, Ren Qi, Cankun Wang,
Hongjun Fu, Qin Ma, and Dong Xu. scgnn is a novel graph neural network framework
[for single-cell rna-seq analyses. Nature communications, 12(1):1–11, 2021. Code Link:](https://github.com/juexinwang/scGNN)
[https://github.com/juexinwang/scGNN.](https://github.com/juexinwang/scGNN)
[415] Kunze Wang, Soyeon Caren Han, Siqu Long, and Josiah Poon. Me-gcn: Multi-dimensional
edge-embedded graph convolutional networks for semi-supervised text classification. arXiv
_preprint arXiv:2204.04618, 2022._
[416] Lei Wang and Ee-Peng Lim. Zero-shot next-item recommendation using large pretrained
language models. arXiv preprint arXiv:2304.03153, 2023.
[417] Minjie Wang, Quan Gan, David Wipf, Zhenkun Cai, Ning Li, Jianheng Tang, Yanlin Zhang,
Zizhao Zhang, Zunyao Mao, Yakun Song, et al. 4dbinfer: A 4d benchmarking toolbox for
graph-centric predictive modeling on relational dbs. arXiv preprint arXiv:2404.18209, 2024.
[418] Qineng Wang, Chen Qian, Xiaochang Li, Ziyu Yao, and Huajie Shao. Lens: A foundation
model for network traffic. arXiv preprint arXiv:2402.03646, 2024.
[419] Song Wang, Yaochen Zhu, Haochen Liu, Zaiyi Zheng, Chen Chen, et al. Knowledge editing
for large language models: A survey. arXiv preprint arXiv:2310.16218, 2023.
[420] Song Wang, Yushun Dong, Binchi Zhang, Zihan Chen, Xingbo Fu, Yinhan He, Cong Shen,
Chuxu Zhang, Nitesh V Chawla, and Jundong Li. Safety in graph machine learning: Threats
and safeguards. arXiv preprint arXiv:2405.11034, 2024.
[421] Tianming Wang, Xiaojun Wan, and Hanqi Jin. Amr-to-text generation with graph transformer.
_Transactions of the Association for Computational Linguistics, 8:19–33, 2020._
[422] Tianyu Wang, Jun Bai, and Sheida Nabavi. Single-cell classification using graph convolutional networks. _BMC Bioinformatics, 22(1):364, 2021._ ISSN 1471-2105. doi: 10.
[1186/s12859-021-04278-2. URL https://bmcbioinformatics.biomedcentral.com/](https://bmcbioinformatics.biomedcentral.com/articles/10.1186/s12859-021-04278-2)
```
articles/10.1186/s12859-021-04278-2.
```
-----
[423] Xiao Wang, Houye Ji, Chuan Shi, Bai Wang, Yanfang Ye, Peng Cui, and Philip S Yu. Heterogeneous graph attention network. In The world wide web conference, pages 2022–2032,
2019.
[424] Xiaoyang Wang, Yao Ma, Yiqi Wang, Wei Jin, Xin Wang, Jiliang Tang, Caiyan Jia, and Jian
Yu. Traffic flow prediction via spatial temporal graph neural network. In Proceedings of the
_web conference 2020, pages 1082–1092, 2020._
[425] Xintao Wang, Qianwen Yang, Yongting Qiu, Jiaqing Liang, Qianyu He, Zhouhong Gu,
Yanghua Xiao, and Wei Wang. Knowledgpt: Enhancing large language models with retrieval
and storage access on knowledge bases. arXiv preprint arXiv:2308.11761, 2023.
[426] Yixin Wang, Jiayuan Ding, Lidan Wu, Aster Wardhani, Patrick Danaher, Qiaolin Lu, Hongzhi
Wen, Wenzhuo Tang, Yi Chang, Yu Leo Lei, et al. Mem-gan: A pseudo membrane generator
for single-cell imaging in fluorescent microscopy. bioRxiv, pages 2023–11, 2023.
[427] Yu Wang, Amin Javari, Janani Balaji, Walid Shalaby, Tyler Derr, and Xiquan Cui. Knowledge
graph-based session recommendation with session-adaptive propagation. In Companion
_Proceedings of the ACM on Web Conference 2024, pages 264–273, 2024._
[428] Yu Wang, Nedim Lipka, Ryan A Rossi, Alexa Siu, Ruiyi Zhang, and Tyler Derr. Knowledge graph prompting for multi-document question answering. In Proceedings of the AAAI
_Conference on Artificial Intelligence, volume 38, pages 19206–19214, 2024._
[429] Yu Wang, Nedim Lipka, Ruiyi Zhang, Alexa Siu, Yuying Zhao, Bo Ni, Xin Wang, Ryan Rossi,
and Tyler Derr. Augmenting textual generation via topology aware retrieval. arXiv preprint
_arXiv:2405.17602, 2024._
[430] Yujie Wang, Hu Zhang, Jiye Liang, and Ru Li. Dynamic heterogeneous-graph reasoning with
language models and knowledge representation learning for commonsense question answering.
In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics
_(Volume 1: Long Papers), pages 14048–14063, 2023._
[431] Yuqi Wang, Boran Jiang, Yi Luo, Dawei He, Peng Cheng, and Liangcai Gao. Reasoning on
efficient knowledge paths: Knowledge graph guides large language model for domain question
answering. arXiv preprint arXiv:2404.10384, 2024.
[432] Zhiruo Wang, Haoyu Dong, Ran Jia, Jia Li, Zhiyi Fu, Shi Han, and Dongmei Zhang. Tuta:
Tree-based transformers for generally structured table pre-training. In Proceedings of the 27th
_ACM SIGKDD Conference on Knowledge Discovery & Data Mining, pages 1780–1790, 2021._
[433] Zichao Wang, Weili Nie, Zhuoran Qiao, Chaowei Xiao, Richard Baraniuk, and Anima Anandkumar. Retrieval-based controllable molecule generation. In The Eleventh International
_Conference on Learning Representations._
[434] Zichao Wang, Weili Nie, Zhuoran Qiao, Chaowei Xiao, Richard Baraniuk, and Anima Anandkumar. Retrieval-based controllable molecule generation. arXiv preprint arXiv:2208.11126,
2022.
[435] Ziqiu Wang, Jun Liu, Shengkai Zhang, and Yang Yang. Poisoned langchain: Jailbreak llms by
[langchain, 2024. URL https://arxiv.org/abs/2406.18122.](https://arxiv.org/abs/2406.18122)
[436] Duncan J Watts and Steven H Strogatz. Collective dynamics of ‘small-world’networks. nature,
393(6684):440–442, 1998.
[437] Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le,
Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models.
_Advances in neural information processing systems, 35:24824–24837, 2022._
[438] Wei Wei, Xubin Ren, Jiabin Tang, Qinyong Wang, Lixin Su, Suqi Cheng, Junfeng Wang,
Dawei Yin, and Chao Huang. Llmrec: Large language models with graph augmentation for
recommendation. In Proceedings of the 17th ACM International Conference on Web Search
_and Data Mining, pages 806–815, 2024._
-----
[439] Yanbin Wei, Qiushi Huang, Yu Zhang, and James Kwok. Kicgpt: Large language model with
knowledge in context for knowledge graph completion. In Findings of the Association for
_Computational Linguistics: EMNLP 2023, pages 8667–8683, 2023._
[440] Hongzhi Wen, Wei Jin, Jiayuan Ding, Christopher Xu, Yuying Xie, and Jiliang Tang. Bichannel masked graph autoencoders for spatially resolved single-cell transcriptomics data
imputation. In NeurIPS 2022 AI for Science: Progress and Promises.
[441] Hongzhi Wen, Jiayuan Ding, Wei Jin, Yiqi Wang, Yuying Xie, and Jiliang Tang. Graph neural
networks for multimodal single-cell data integration. In Proceedings of the 28th ACM SIGKDD
_conference on knowledge discovery and data mining, pages 4153–4163, 2022._
[442] Hongzhi Wen, Wenzhuo Tang, Wei Jin, Jiayuan Ding, Renming Liu, Feng Shi, Yuying Xie,
and Jiliang Tang. Single cells are spatial tokens: Transformers for spatial transcriptomic data
imputation. arXiv preprint arXiv:2302.03038, 2023.
[443] Yilin Wen, Zifeng Wang, and Jimeng Sun. Mindmap: Knowledge graph prompting sparks
graph of thoughts in large language models. arXiv preprint arXiv:2308.09729, 2023.
[444] Tim Weninger, Yonatan Bisk, and Jiawei Han. Document-topic hierarchies from document
graphs. In Proceedings of the 21st ACM international conference on Information and knowl_edge management, pages 635–644, 2012._
[445] Daniel S Wigh, Jonathan M Goodman, and Alexei A Lapkin. A review of molecular representation in the age of machine learning. Wiley Interdisciplinary Reviews: Computational
_Molecular Science, 12(5):e1603, 2022._
[446] Nirmalie Wiratunga, Ramitha Abeyratne, Lasal Jayawardena, Kyle Martin, Stewart Massie,
Ikechukwu Nkisi-Orji, Ruvan Weerasinghe, Anne Liret, and Bruno Fleisch. Cbr-rag: Casebased reasoning for retrieval augmented generation in llms for legal question answering. In
_International Conference on Case-Based Reasoning, pages 445–460. Springer, 2024._
[447] Eugene Wong and Karel Youssefi. Decomposition—a strategy for query processing. ACM
_Transactions on Database Systems (TODS), 1(3):223–241, 1976._
[448] Bo Wu, Bo Lang, and Yang Liu. Gksh: Graph based image retrieval using supervised kernel
hashing. In Proceedings of the International Conference on Internet Multimedia Computing
_and Service, pages 88–93, 2016._
[449] Junde Wu, Jiayuan Zhu, and Yunli Qi. Medical graph rag: Towards safe medical large language
model via graph retrieval-augmented generation. arXiv preprint arXiv:2408.04187, 2024.
[450] Lingfei Wu, Yu Chen, Kai Shen, Xiaojie Guo, Hanning Gao, Shucheng Li, Jian Pei, Bo Long,
et al. Graph neural networks for natural language processing: A survey. Foundations and
_Trends® in Machine Learning, 16(2):119–328, 2023._
[451] Qitian Wu, Chenxiao Yang, and Junchi Yan. Towards open-world feature extrapolation: An
inductive graph learning approach. Advances in Neural Information Processing Systems, 34:
19435–19447, 2021.
[452] Shirley Wu, Shiyu Zhao, Michihiro Yasunaga, Kexin Huang, Kaidi Cao, Qian Huang, Vassilis N Ioannidis, Karthik Subbian, James Zou, and Jure Leskovec. Stark: Benchmarking llm
retrieval on textual and relational knowledge bases. arXiv preprint arXiv:2404.13207, 2024.
[453] Xiaobao Wu, Liangming Pan, William Yang Wang, and Anh Tuan Luu. Updating language models with unstructured facts: Towards practical knowledge editing. arXiv preprint
_arXiv:2402.18909, 2024._
[454] Xixi Wu, Yifei Shen, Caihua Shan, Kaitao Song, Siwei Wang, Bohang Zhang, Jiarui Feng,
Hong Cheng, Wei Chen, Yun Xiong, et al. Can graph learning improve task planning? arXiv
_preprint arXiv:2405.19119, 2024._
[455] Yike Wu, Nan Hu, Sheng Bi, Guilin Qi, Jie Ren, Anhuan Xie, and Wei Song. Retrieve-rewriteanswer: A kg-to-text enhanced llms framework for knowledge graph question answering.
_arXiv preprint arXiv:2309.11206, 2023._
-----
[456] Yike Wu, Nan Hu, Guilin Qi, Sheng Bi, Jie Ren, Anhuan Xie, and Wei Song. Retrieve-rewriteanswer: A kg-to-text enhanced llms framework for knowledge graph question answering.
_arXiv preprint arXiv:2309.11206, 2023._
[457] Yike Wu, Yi Huang, Nan Hu, Yuncheng Hua, Guilin Qi, Jiaoyan Chen, and Jeff Z Pan. Cotkr:
Chain-of-thought enhanced knowledge rewriting for complex knowledge graph question
answering. arXiv preprint arXiv:2409.19753, 2024.
[458] Yunjia Xi, Weiwen Liu, Jianghao Lin, Xiaoling Cai, Hong Zhu, Jieming Zhu, Bo Chen,
Ruiming Tang, Weinan Zhang, and Yong Yu. Towards open-world recommendation with
knowledge augmentation from large language models. In Proceedings of the 18th ACM
_Conference on Recommender Systems, pages 12–22, 2024._
[459] Yu Xia, Junda Wu, Sungchul Kim, Tong Yu, Ryan A Rossi, Haoliang Wang, and Julian
McAuley. Knowledge-aware query expansion with large language models for textual and
relational retrieval. arXiv preprint arXiv:2410.13765, 2024.
[460] Chong Xiang, Tong Wu, Zexuan Zhong, David Wagner, Danqi Chen, and Prateek Mittal.
Certifiably robust rag against retrieval corruption. arXiv preprint arXiv:2405.15556, 2024.
[461] Shunxin Xiao, Shiping Wang, Yuanfei Dai, and Wenzhong Guo. Graph neural networks in
node classification: survey and evaluation. Machine Vision and Applications, 33(1):4, 2022.
[462] Qianqian Xie, Jimin Huang, Tulika Saha, and Sophia Ananiadou. Gretel: Graph contrastive
topic enhanced language model for long document extractive summarization. arXiv preprint
_arXiv:2208.09982, 2022._
[463] Zhouhang Xie, Sameer Singh, Julian McAuley, and Bodhisattwa Prasad Majumder. Factual
and informative review generation for explainable recommendation. In Proceedings of the
_AAAI Conference on Artificial Intelligence, volume 37, pages 13816–13824, 2023._
[464] Amy Xin, Yunjia Qi, Zijun Yao, Fangwei Zhu, Kaisheng Zeng, Xu Bin, Lei Hou, and Juanzi
Li. Llmael: Large language models are good context augmenters for entity linking. arXiv
_preprint arXiv:2407.04020, 2024._
[465] Lu Xing and Lina Sela. Graph neural networks for state estimation in water distribution
systems: Application of supervised and semisupervised learning. Journal of Water Resources
_Planning and Management, 148(5):04022018, 2022._
[466] Guangzhi Xiong, Qiao Jin, Zhiyong Lu, and Aidong Zhang. Benchmarking retrievalaugmented generation for medicine. arXiv preprint arXiv:2312.10997, 2023.
[467] Guangzhi Xiong, Qiao Jin, Zhiyong Lu, and Aidong Zhang. Benchmarking retrievalaugmented generation for medicine. arXiv preprint arXiv:2402.13178, 2024.
[468] Wenhan Xiong, Xiang Lorraine Li, Srini Iyer, Jingfei Du, Patrick Lewis, William Yang Wang,
Yashar Mehdad, Wen-tau Yih, Sebastian Riedel, Douwe Kiela, et al. Answering complex
open-domain questions with multi-hop dense retrieval. arXiv preprint arXiv:2009.12756,
2020.
[469] Chenyang Xu, Lei Cai, and Jingyang Gao. An efficient scrna-seq dropout imputation method
using graph attention network. BMC bioinformatics, 22:1–18, 2021.
[470] Kun Xu, Lingfei Wu, Zhiguo Wang, Yansong Feng, Michael Witbrock, and Vadim Sheinin.
Graph2seq: Graph to sequence learning with attention-based neural networks. arXiv preprint
_arXiv:1804.00823, 2018._
[471] Mingzhou Xu, Liangyou Li, Derek Wong, Qun Liu, Lidia S Chao, et al. Document graph for
neural machine translation. arXiv preprint arXiv:2012.03477, 2020.
[472] Ran Xu, Wenqi Shi, Yue Yu, Yuchen Zhuang, Bowen Jin, May D Wang, Joyce C Ho, and Carl
Yang. Ram-ehr: Retrieval augmentation meets clinical predictions on electronic health records.
_arXiv preprint arXiv:2403.00815, 2024._
-----
[473] Rongwu Xu, Zehan Qi, Zhijiang Guo, Cunxiang Wang, Hongru Wang, Yue Zhang, and Wei
Xu. Knowledge conflicts for llms: A survey. arXiv preprint arXiv:2403.08319, 2024.
[474] Shicheng Xu, Liang Pang, Huawei Shen, and Xueqi Cheng. Unveil the duality of
retrieval-augmented generation: Theoretical analysis and practical solution. arXiv preprint
_arXiv:2406.00944, 2024._
[475] Weiwen Xu, Yang Deng, Huihui Zhang, Deng Cai, and Wai Lam. Exploiting reasoning chains
for multi-hop science question answering. arXiv preprint arXiv:2109.02905, 2021.
[476] Weiwen Xu, Huihui Zhang, Deng Cai, and Wai Lam. Dynamic semantic graph construction and reasoning for explainable multi-hop science question answering. arXiv preprint
_arXiv:2105.11776, 2021._
[477] Weiye Xu, Min Wang, Wengang Zhou, and Houqiang Li. P-rag: Progressive retrieval augmented generation for planning on embodied everyday task. In ACM Multimedia 2024.
[478] Jiaqi Xue, Mengxin Zheng, Yebowen Hu, Fei Liu, Xun Chen, and Qian Lou. Badrag:
Identifying vulnerabilities in retrieval augmented generation of large language models. arXiv
_preprint arXiv:2406.00083, 2024._
[479] Wenyuan Xue, Baosheng Yu, Wen Wang, Dacheng Tao, and Qingyong Li. Tgrnet: A table
graph reconstruction network for table structure recognition. In Proceedings of the IEEE/CVF
_International Conference on Computer Vision, pages 1295–1304, 2021._
[480] Jiahuan Yan, Jintai Chen, Yixuan Wu, Danny Z Chen, and Jian Wu. T2g-former: organizing
tabular features into relation graphs promotes heterogeneous feature interaction. In Proceedings
_of the AAAI Conference on Artificial Intelligence, volume 37, pages 10720–10728, 2023._
[481] Diji Yang, Jinmeng Rao, Kezhen Chen, Xiaoyuan Guo, Yawen Zhang, Jie Yang, and Yi Zhang.
Im-rag: Multi-round retrieval-augmented generation through learning inner monologues. In
_Proceedings of the 47th International ACM SIGIR Conference on Research and Development_
_in Information Retrieval, pages 730–740, 2024._
[482] Kang Yang, Kamal Al-Sabahi, Yanmin Xiang, and Zuping Zhang. An integrated graph model
for document summarization. Information, 9(9):232, 2018.
[483] Rui Yang, Haoran Liu, Qingcheng Zeng, Yu He Ke, Wanxin Li, Lechao Cheng, Qingyu Chen,
James Caverlee, Yutaka Matsuo, and Irene Li. Kg-rank: Enhancing large language models for
medical qa with knowledge graphs and ranking techniques. arXiv preprint arXiv:2403.05881,
2024.
[484] Rui Yang, Haoran Liu, Qingcheng Zeng, Yu He Ke, Wanxin Li, Lechao Cheng, Qingyu Chen,
James Caverlee, Yutaka Matsuo, and Irene Li. Kg-rank: Enhancing large language models for
medical qa with knowledge graphs and ranking techniques. arXiv preprint arXiv:2403.05881,
2024.
[485] Shenghao Yang, Weizhi Ma, Peijie Sun, Min Zhang, Qingyao Ai, Yiqun Liu, and Mingchen
Cai. Common sense enhanced knowledge-based recommendation with large language model.
_arXiv preprint arXiv:2403.18325, 2024._
[486] Zhaoming Yang, Zhe Liu, Jing Zhou, Chaofan Song, Qi Xiang, Qian He, Jingjing Hu,
Michael H Faber, Enrico Zio, Zhenlin Li, et al. A graph neural network (gnn) method
for assigning gas calorific values to natural gas pipeline networks. Energy, 278:127875, 2023.
[487] Zukang Yang and Zixuan Zhu. Curiousllm: Elevating multi-document qa with reasoninginfused knowledge graph prompting. arXiv preprint arXiv:2404.09077, 2024.
[488] Shunyu Yao, Dian Yu, Jeffrey Zhao, Izhak Shafran, Tom Griffiths, Yuan Cao, and Karthik
Narasimhan. Tree of thoughts: Deliberate problem solving with large language models.
_Advances in Neural Information Processing Systems, 36, 2024._
[489] Yifan Yao, Jinhao Duan, Kaidi Xu, Yuanfang Cai, Zhibo Sun, and Yue Zhang. A survey on large
language model (llm) security and privacy: The good, the bad, and the ugly. High-Confidence
_Computing, page 100211, 2024._
-----
[490] Zijun Yao, Weijian Qi, Liangming Pan, Shulin Cao, Linmei Hu, Weichuan Liu, Lei Hou, and
Juanzi Li. Seakr: Self-aware knowledge retrieval for adaptive retrieval augmented generation.
_arXiv preprint arXiv:2406.19215, 2024._
[491] Michihiro Yasunaga, Rui Zhang, Kshitijh Meelu, Ayush Pareek, Krishnan Srinivasan, and
Dragomir Radev. Graph-based neural multi-document summarization. _arXiv preprint_
_arXiv:1706.06681, 2017._
[492] Michihiro Yasunaga, Hongyu Ren, Antoine Bosselut, Percy Liang, and Jure Leskovec. Qagnn: Reasoning with language models and knowledge graphs for question answering. In
_Proceedings of the 2021 Conference of the North American Chapter of the Association for_
_Computational Linguistics: Human Language Technologies, pages 535–546, 2021._
[493] Michihiro Yasunaga, Antoine Bosselut, Hongyu Ren, Xikun Zhang, Christopher D Manning,
Percy S Liang, and Jure Leskovec. Deep bidirectional language-knowledge graph pretraining.
_Advances in Neural Information Processing Systems, 35:37309–37323, 2022._
[494] Michihiro Yasunaga, Jure Leskovec, and Percy Liang. Linkbert: Pretraining language models
with document links. arXiv preprint arXiv:2203.15827, 2022.
[495] Fanghua Ye, Mingming Yang, Jianhui Pang, Longyue Wang, Derek F. Wong, Emine Yilmaz,
Shuming Shi, and Zhaopeng Tu. Benchmarking llms via uncertainty quantification, 2024.
[URL https://arxiv.org/abs/2401.12794.](https://arxiv.org/abs/2401.12794)
[496] Ori Yoran, Tomer Wolfson, Ori Ram, and Jonathan Berant. Making retrieval-augmented
language models robust to irrelevant context. arXiv preprint arXiv:2310.01558, 2023.
[497] Jiaxuan You, Xiaobai Ma, Yi Ding, Mykel J Kochenderfer, and Jure Leskovec. Handling
missing data with graph representation learning. Advances in Neural Information Processing
_Systems, 33:19075–19087, 2020._
[498] Donghan Yu, Chenguang Zhu, Yuwei Fang, Wenhao Yu, Shuohang Wang, Yichong Xu, Xiang
Ren, Yiming Yang, and Michael Zeng. Kg-fid: Infusing knowledge graph in fusion-in-decoder
for open-domain question answering. arXiv preprint arXiv:2110.04330, 2021.
[499] Junchi Yu, Ran He, and Rex Ying. Thought propagation: An analogical approach to complex
reasoning with large language models. arXiv preprint arXiv:2310.03965, 2023.
[500] Xinli Yu, Zheng Chen, Yuan Ling, Shujing Dong, Zongyi Liu, and Yanbin Lu. Temporal data
meets llm–explainable financial time series forecasting. arXiv preprint arXiv:2306.11025,
2023.
[501] Xueli Yu, Weizhi Xu, Zeyu Cui, Shu Wu, and Liang Wang. Graph-based hierarchical relevance
matching signals for ad-hoc retrieval. In Proceedings of the Web Conference 2021, pages
778–787, 2021.
[502] Zhuohan Yu, Yifu Lu, Yunhe Wang, Fan Tang, Ka-Chun Wong, and Xiangtao Li. Zinb-based
graph embedding autoencoder for single-cell rna-seq interpretations. In Proceedings of the
_AAAI conference on artificial intelligence, volume 36, pages 4671–4679, 2022._
[503] Siyu Yuan, Jiangjie Chen, Changzhi Sun, Jiaqing Liang, Yanghua Xiao, and Deqing Yang.
Analogykb: Unlocking analogical reasoning of language models with a million-scale knowledge base. arXiv preprint arXiv:2305.05994, 2023.
[504] Seongjun Yun, Minbyul Jeong, Raehyun Kim, Jaewoo Kang, and Hyunwoo J Kim. Graph
transformer networks. Advances in neural information processing systems, 32, 2019.
[505] Ahtsham Zafar, Venkatesh Balavadhani Parthasarathy, Chan Le Van, Saad Shahid, Arsalan
Shahid, et al. Building trust in conversational ai: A comprehensive review and solution
architecture for explainable, privacy-aware systems using llms and knowledge graph. arXiv
_preprint arXiv:2308.13534, 2023._
[506] Lukáš Zahradník, Jan Neumann, and Gustav Šír. A deep learning blueprint for relational
databases. In NeurIPS 2023 Second Table Representation Learning Workshop, 2023.
-----
[507] Cyril Zakka, Rohan Shad, Akash Chaurasia, Alex R Dalal, Jennifer L Kim, Michael Moor,
Robyn Fong, Curran Phillips, Kevin Alexander, Euan Ashley, et al. Almanac—retrievalaugmented language models for clinical medicine. NEJM AI, 1(2):AIoa2300068, 2024.
[508] Xuan Zang, Xianbing Zhao, and Buzhou Tang. Hierarchical molecular graph self-supervised
learning for property prediction. Communications Chemistry, 6(1):34, 2023.
[509] Deborah A Zarin, Tony Tse, Rebecca J Williams, Robert M Califf, and Nicholas C Ide. The
clinicaltrials. gov results database—update and key issues. New England Journal of Medicine,
364(9):852–860, 2011.
[510] Huimin Zeng, Zhenrui Yue, Qian Jiang, and Dong Wang. Federated recommendation via
hybrid retrieval augmented generation. arXiv preprint arXiv:2403.04256, 2024.
[511] Jingying Zeng, Richard Huang, Waleed Malik, Langxuan Yin, Bojan Babic, Danny Shacham,
Xiao Yan, Jaewon Yang, and Qi He. Large language models for social networks: Applications,
challenges, and solutions. arXiv preprint arXiv:2401.02575, 2024.
[512] Shenglai Zeng, Jiankun Zhang, Pengfei He, Jie Ren, Tianqi Zheng, Hanqing Lu, Han Xu,
Hui Liu, Yue Xing, and Jiliang Tang. Mitigating the privacy issues in retrieval-augmented
generation (rag) via pure synthetic data. arXiv preprint arXiv:2406.14773, 2024.
[513] Shenglai Zeng, Jiankun Zhang, Pengfei He, Yue Xing, Yiding Liu, Han Xu, Jie Ren, Shuaiqiang
Wang, Dawei Yin, Yi Chang, et al. The good and the bad: Exploring privacy issues in retrievalaugmented generation (rag). arXiv preprint arXiv:2402.16893, 2024.
[514] Shenglai Zeng, Jiankun Zhang, Bingheng Li, Yuping Lin, Tianqi Zheng, Dante Everaert, Hanqing Lu, Hui Liu, Yue Xing, Monica Xiao Cheng, et al. Towards knowledge
checking in retrieval-augmented generation: A representation perspective. arXiv preprint
_arXiv:2411.14572, 2024._
[515] Boyu Zhang, Hongyang Yang, Tianyu Zhou, Muhammad Ali Babar, and Xiao-Yang Liu.
Enhancing financial sentiment analysis via retrieval augmented large language models. In
_Proceedings of the fourth ACM international conference on AI in finance, pages 349–356,_
2023.
[516] Chen Zhang, Qiuchi Li, and Dawei Song. Aspect-based sentiment classification with aspectspecific graph convolutional networks. arXiv preprint arXiv:1909.03477, 2019.
[517] Han Zhang, Quan Gan, David Wipf, and Weinan Zhang. Gfs: Graph-based feature synthesis
for prediction over relational database. Proceedings of the VLDB Endowment. ISSN, 2150:
8097, 2023.
[518] Haopeng Zhang and Jiawei Zhang. Text graph transformer for document classification. In
_Conference on empirical methods in natural language processing (EMNLP), 2020._
[519] Haopeng Zhang, Xiao Liu, and Jiawei Zhang. Contrastive hierarchical discourse graph for
scientific document summarization. In The 4th workshop on Computa-tional Approaches to
_Discourse at the 61st Annual Meeting of the Association for Computational Linguistics (ACL_
_CODI’23), July 9-14, 2023. Toronto, Canada., 2023._
[520] Haozhen Zhang, Tao Feng, and Jiaxuan You. Graph of records: Boosting retrieval augmented
generation for long-context summarization with graphs. arXiv preprint arXiv:2410.11001,
2024.
[521] Jing Zhang, Biao Liu, Jie Tang, Ting Chen, and Juanzi Li. Social influence locality for
modeling retweeting behaviors. In Twenty-third international joint conference on artificial
_intelligence. Citeseer, 2013._
[522] Jing Zhang, Xiaokang Zhang, Jifan Yu, Jian Tang, Jie Tang, Cuiping Li, and Hong Chen.
Subgraph retrieval enhanced model for multi-hop knowledge base question answering. In
_Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics_
_(Volume 1: Long Papers), pages 5773–5784, 2022._
-----
[523] Jintian Zhang, Xin Xu, Ningyu Zhang, Ruibo Liu, Bryan Hooi, and Shumin Deng. Exploring collaboration mechanisms for llm agents: A social psychology view. arXiv preprint
_arXiv:2310.02124, 2023._
[524] Lihui Zhang and Ruifan Li. Ke-gcl: Knowledge enhanced graph contrastive learning for commonsense question answering. In Findings of the Association for Computational Linguistics:
_EMNLP 2022, pages 76–87, 2022._
[525] Ming Zhang, Jiyu Lu, Jiahao Yang, Jun Zhou, Meilin Wan, and Xuejun Zhang. From coarse
to fine: Enhancing multi-document summarization with multi-granularity relationship-based
extractor. Information Processing & Management, 61(3):103696, 2024.
[526] Ningyu Zhang, Xiang Chen, Xin Xie, Shumin Deng, Chuanqi Tan, Mosha Chen, Fei Huang,
Luo Si, and Huajun Chen. Document-level relation extraction as semantic segmentation. arXiv
_preprint arXiv:2106.03618, 2021._
[527] Shichang Zhang, Da Zheng, Jiani Zhang, Qi Zhu, Soji Adeshina, Christos Faloutsos, George
Karypis, Yizhou Sun, et al. Hierarchical compression of text-rich graphs via large language
models. arXiv preprint arXiv:2406.11884, 2024.
[528] Wei Zhang, Zeyuan Chen, Hongyuan Zha, and Jianyong Wang. Learning from substitutable
and complementary relations for graph-based sequential product recommendation. ACM
_Transactions on Information Systems (TOIS), 40(2):1–28, 2021._
[529] Wen Zhang, Yanbin Lu, Bella Dubrov, Zhi Xu, Shang Shang, and Emilio Maldonado. Deep
hierarchical product classification based on pre-trained multilingual knowledge. 2021.
[530] Xikun Zhang, Antoine Bosselut, Michihiro Yasunaga, Hongyu Ren, Percy Liang, Christopher D Manning, and Jure Leskovec. Greaselm: Graph reasoning enhanced language models.
In International Conference on Learning Representations.
[531] Xingyao Zhang, Linjun Shou, Jian Pei, Ming Gong, Lijie Wen, and Daxin Jiang. A graph representation of semi-structured data for web question answering. arXiv preprint arXiv:2010.06801,
2020.
[532] Xuanyu Zhang. Cfgnn: Cross flow graph neural networks for question answering on complex
tables. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 34, pages
9596–9603, 2020.
[533] Xulang Zhang, Rui Mao, and Erik Cambria. A survey on syntactic processing techniques.
_Artificial Intelligence Review, 56(6):5645–5728, 2023._
[534] Yi Zhang, Yuying Zhao, Zhaoqing Li, Xueqi Cheng, Yu Wang, Olivera Kotevska, S Yu Philip,
and Tyler Derr. A survey on privacy in graph neural networks: Attacks, preservation, and
applications. IEEE Transactions on Knowledge and Data Engineering, 2024.
[535] Yichi Zhang, Zhuo Chen, Wen Zhang, and Huajun Chen. Making large language models
perform better in knowledge graph completion. arXiv preprint arXiv:2310.06671, 2023.
[536] Yu Zhang, Hao Cheng, Zhihong Shen, Xiaodong Liu, Ye-Yi Wang, and Jianfeng Gao. Pretraining multi-task contrastive learning models for scientific literature understanding. arXiv
_preprint arXiv:2305.14232, 2023._
[537] Yue Zhang, Zhihao Zhang, Wenbin Lai, Chong Zhang, Tao Gui, Qi Zhang, and Xuan-Jing
Huang. to-tree: Parsing pdf text blocks into a tree. In Findings of the Association for
_Computational Linguistics: EMNLP 2024, pages 10704–10714, 2024._
[538] Yufeng Zhang, Xueli Yu, Zeyu Cui, Shu Wu, Zhongzhen Wen, and Liang Wang. Every
document owns its structure: Inductive text classification via graph neural networks. arXiv
_preprint arXiv:2004.13826, 2020._
[539] Yuhao Zhang, Peng Qi, and Christopher D Manning. Graph convolution over pruned dependency trees improves relation extraction. arXiv preprint arXiv:1809.10185, 2018.
-----
[540] Yunyi Zhang, Ruozhen Yang, Xueqiang Xu, Jinfeng Xiao, Jiaming Shen, and Jiawei Han. Teleclass: Taxonomy enrichment and llm-enhanced hierarchical text classification with minimal
supervision. arXiv preprint arXiv:2403.00165, 2024.
[541] Yuyu Zhang, Ping Nie, Arun Ramamurthy, and Le Song. Answering any-hop open-domain
questions with iterative document reranking. In Proceedings of the 44th International ACM
_SIGIR Conference on Research and Development in Information Retrieval, pages 481–490,_
2021.
[542] Zhehao Zhang, Jiaao Chen, and Diyi Yang. Darg: Dynamic evaluation of large language
models via adaptive reasoning graph. arXiv preprint arXiv:2406.17271, 2024.
[543] Zhenyu Zhang, Bowen Yu, Xiaobo Shu, Xue Mengge, Tingwen Liu, and Li Guo. From what
to why: Improving relation extraction with rationale graph. In Findings of the Association for
_Computational Linguistics: ACL-IJCNLP 2021, pages 86–95, 2021._
[544] Zhiqiang Zhang, Linan Wang, Xiaoqin Xie, and Haiwei Pan. A graph based document retrieval
method. In 2018 IEEE 22nd International Conference on Computer Supported Cooperative
_Work in Design ((CSCWD)), pages 426–432. IEEE, 2018._
[545] Ziwei Zhang, Peng Cui, and Wenwu Zhu. Deep learning on graphs: A survey. _IEEE_
_Transactions on Knowledge and Data Engineering, 34(1):249–270, 2020._
[546] Zixuan Zhang and Heng Ji. Abstract meaning representation guided graph encoding and
decoding for joint information extraction. In Proc. The 2021 Conference of the North American
_Chapter of the Association for Computational Linguistics-Human Language Technologies_
_(NAACL-HLT2021), 2021._
[547] Zixuan Zhang, Heba Elfardy, Markus Dreyer, Kevin Small, Heng Ji, and Mohit Bansal.
Enhancing multi-document summarization with cross-document graph-based information
extraction. In Proceedings of the 17th Conference of the European Chapter of the Association
_for Computational Linguistics, pages 1696–1707, 2023._
[548] Chenlong Zhao, Xiwen Zhou, Xiaopeng Xie, and Yong Zhang. Hierarchical attention graph for
scientific document summarization in global and local level. arXiv preprint arXiv:2405.10202,
2024.
[549] Haiteng Zhao, Shengchao Liu, Ma Chang, Hannan Xu, Jie Fu, Zhihong Deng, Lingpeng Kong,
and Qi Liu. Gimlet: A unified graph-text model for instruction-based molecule zero-shot
learning. Advances in Neural Information Processing Systems, 36:5850–5887, 2023.
[550] Haiyan Zhao, Hanjie Chen, Fan Yang, Ninghao Liu, Huiqi Deng, Hengyi Cai, Shuaiqiang
Wang, Dawei Yin, and Mengnan Du. Explainability for large language models: A survey.
_ACM Transactions on Intelligent Systems and Technology, 15(2):1–38, 2024._
[551] Penghao Zhao, Hailin Zhang, Qinhan Yu, Zhengren Wang, Yunteng Geng, Fangcheng Fu,
Ling Yang, Wentao Zhang, and Bin Cui. Retrieval-augmented generation for ai-generated
content: A survey. arXiv preprint arXiv:2402.19473, 2024.
[552] Qifang Zhao, Weidong Ren, Tianyu Li, Xiaoxiao Xu, and Hong Liu. Graphgpt: Graph learning
with generative pre-trained transformers. arXiv preprint arXiv:2401.00529, 2023.
[553] Wayne Xin Zhao, Shanlei Mu, Yupeng Hou, Zihan Lin, Yushuo Chen, Xingyu Pan, Kaiyuan
Li, Yujie Lu, Hui Wang, Changxin Tian, et al. Recbole: Towards a unified, comprehensive
and efficient framework for recommendation algorithms. In proceedings of the 30th acm
_international conference on information & knowledge management, pages 4653–4664, 2021._
[554] Zihuai Zhao, Wenqi Fan, Jiatong Li, Yunqing Liu, Xiaowei Mei, Yiqi Wang, Zhen Wen, Fei
Wang, Xiangyu Zhao, Jiliang Tang, et al. Recommender systems in the era of large language
models (llms). IEEE Transactions on Knowledge and Data Engineering, 2024.
[555] Zirui Zhao, Wee Sun Lee, and David Hsu. Large language models as commonsense knowledge
for large-scale task planning. Advances in Neural Information Processing Systems, 36, 2024.
-----
[556] Chen Zheng and Parisa Kordjamshidi. Srlgrn: Semantic role labeling graph reasoning network.
_arXiv preprint arXiv:2010.03604, 2020._
[557] Jiajun Zhong, Ning Gui, and Weiwei Ye. Data imputation with iterative graph reconstruction. In
_Proceedings of the AAAI Conference on Artificial Intelligence, volume 37, pages 11399–11407,_
2023.
[558] Huiwei Zhou, Yibin Xu, Weihong Yao, Zhe Liu, Chengkun Lang, and Haibin Jiang. Global
context-enhanced graph convolutional networks for document-level relation extraction. In
_Proceedings of the 28th International Conference on Computational Linguistics, pages 5259–_
5270, 2020.
[559] Kaixiong Zhou, Zirui Liu, Rui Chen, Li Li, Soo-Hyun Choi, and Xia Hu. Table2graph:
Transforming tabular data to unified weighted graph. In IJCAI, pages 2420–2426, 2022.
[560] Qihang Zhou, Jiming Chen, Haoyu Liu, Shibo He, and Wenchao Meng. Detecting multivariate
time series anomalies with zero known label. In Proceedings of the AAAI Conference on
_Artificial Intelligence, volume 37, pages 4963–4971, 2023._
[561] Yujia Zhou, Yan Liu, Xiaoxi Li, Jiajie Jin, Hongjin Qian, Zheng Liu, Chaozhuo Li, Zhicheng
Dou, Tsung-Yi Ho, and Philip S Yu. Trustworthiness in retrieval-augmented generation
systems: A survey. arXiv preprint arXiv:2409.10102, 2024.
[562] Fengbin Zhu, Wenqiang Lei, Chao Wang, Jianming Zheng, Soujanya Poria, and Tat-Seng
Chua. Retrieving and reading: A comprehensive survey on open-domain question answering.
_arXiv preprint arXiv:2101.00774, 2021._
[563] Jiong Zhu, Yujun Yan, Lingxiao Zhao, Mark Heimann, Leman Akoglu, and Danai Koutra. Beyond homophily in graph neural networks: Current limitations and effective designs. Advances
_in neural information processing systems, 33:7793–7804, 2020._
[564] Kaijie Zhu, Jindong Wang, Qinlin Zhao, Ruochen Xu, and Xing Xie. Dyval 2: Dynamic
evaluation of large language models by meta probing agents. arXiv preprint arXiv:2402.14865,
2024.
[565] Lixing Zhu, Yulan He, and Deyu Zhou. Neural opinion dynamics model for the prediction of
user-level stance dynamics. Information Processing & Management, 57(2):102031, 2020.
[566] Yinghao Zhu, Changyu Ren, Zixiang Wang, Xiaochen Zheng, Shiyun Xie, Junlan Feng,
Xi Zhu, Zhoujun Li, Liantao Ma, and Chengwei Pan. Emerge: Integrating rag for improved
multimodal ehr predictive modeling. arXiv preprint arXiv:2406.00036, 2024.
[567] Yinghao Zhu, Changyu Ren, Shiyun Xie, Shukai Liu, Hangyuan Ji, Zixiang Wang, Tao Sun,
Long He, Zhoujun Li, Xi Zhu, et al. Realm: Rag-driven enhancement of multimodal electronic
health records analysis via large language models. arXiv preprint arXiv:2402.07016, 2024.
[568] Yuqi Zhu, Shuofei Qiao, Yixin Ou, Shumin Deng, Ningyu Zhang, Shiwei Lyu, Yue Shen,
Lei Liang, Jinjie Gu, and Huajun Chen. Knowagent: Knowledge-augmented planning for
llm-based agents. arXiv preprint arXiv:2403.03101, 2024.
[569] Yuchen Zhuang, Xiang Chen, Tong Yu, Saayan Mitra, Victor Bursztyn, Ryan A Rossi, Somdeb
Sarkhel, and Chao Zhang. Toolchain*: Efficient action space navigation in large language
models with a* search. In The Twelfth International Conference on Learning Representations.
[570] Yuchen Zhuang, Xiang Chen, Tong Yu, Saayan Mitra, Victor Bursztyn, Ryan A Rossi, Somdeb
Sarkhel, and Chao Zhang. Toolchain*: Efficient action space navigation in large language
models with a* search. arXiv preprint arXiv:2310.13227, 2023.
[571] Ray Daniel Zimmerman, Carlos Edmundo Murillo-Sánchez, and Robert John Thomas. Matpower: Steady-state operations, planning, and analysis tools for power systems research and
education. IEEE Transactions on power systems, 26(1):12–19, 2010.
[572] Tao Zou, Le Yu, Yifei Huang, Leilei Sun, and Bowen Du. Pretraining language models with
text-attributed heterogeneous graphs. arXiv preprint arXiv:2310.12580, 2023.
-----
[573] Wei Zou, Runpeng Geng, Binghui Wang, and Jinyuan Jia. Poisonedrag: Knowledge corruption
[attacks to retrieval-augmented generation of large language models, 2024. URL https:](https://arxiv.org/abs/2402.07867)
```
//arxiv.org/abs/2402.07867.
```
-----
|
{
"id": "2501.00309",
"submitter": "Yu Wang",
"authors": "Haoyu Han, Yu Wang, Harry Shomer, Kai Guo, Jiayuan Ding, Yongjia Lei,\n Mahantesh Halappanavar, Ryan A. Rossi, Subhabrata Mukherjee, Xianfeng Tang,\n Qi He, Zhigang Hua, Bo Long, Tong Zhao, Neil Shah, Amin Javari, Yinglong Xia,\n Jiliang Tang",
"title": "Retrieval-Augmented Generation with Graphs (GraphRAG)",
"comments": null,
"journal-ref": null,
"doi": null,
"report-no": null,
"categories": "cs.IR cs.CL cs.LG",
"license": "http://arxiv.org/licenses/nonexclusive-distrib/1.0/",
"abstract": " Retrieval-augmented generation (RAG) is a powerful technique that enhances\ndownstream task execution by retrieving additional information, such as\nknowledge, skills, and tools from external sources. Graph, by its intrinsic\n\"nodes connected by edges\" nature, encodes massive heterogeneous and relational\ninformation, making it a golden resource for RAG in tremendous real-world\napplications. As a result, we have recently witnessed increasing attention on\nequipping RAG with Graph, i.e., GraphRAG. However, unlike conventional RAG,\nwhere the retriever, generator, and external data sources can be uniformly\ndesigned in the neural-embedding space, the uniqueness of graph-structured\ndata, such as diverse-formatted and domain-specific relational knowledge, poses\nunique and significant challenges when designing GraphRAG for different\ndomains. Given the broad applicability, the associated design challenges, and\nthe recent surge in GraphRAG, a systematic and up-to-date survey of its key\nconcepts and techniques is urgently desired. Following this motivation, we\npresent a comprehensive and up-to-date survey on GraphRAG. Our survey first\nproposes a holistic GraphRAG framework by defining its key components,\nincluding query processor, retriever, organizer, generator, and data source.\nFurthermore, recognizing that graphs in different domains exhibit distinct\nrelational patterns and require dedicated designs, we review GraphRAG\ntechniques uniquely tailored to each domain. Finally, we discuss research\nchallenges and brainstorm directions to inspire cross-disciplinary\nopportunities. Our survey repository is publicly maintained at\nhttps://github.com/Graph-RAG/GraphRAG/.\n",
"versions": {
"version": [
"v1",
"v2"
],
"created": [
"Tue, 31 Dec 2024 06:59:35 GMT",
"Wed, 8 Jan 2025 05:16:25 GMT"
]
},
"update_date": "2025-01-09",
"authors_parsed": [
[
"Han",
"Haoyu",
""
],
[
"Wang",
"Yu",
""
],
[
"Shomer",
"Harry",
""
],
[
"Guo",
"Kai",
""
],
[
"Ding",
"Jiayuan",
""
],
[
"Lei",
"Yongjia",
""
],
[
"Halappanavar",
"Mahantesh",
""
],
[
"Rossi",
"Ryan A.",
""
],
[
"Mukherjee",
"Subhabrata",
""
],
[
"Tang",
"Xianfeng",
""
],
[
"He",
"Qi",
""
],
[
"Hua",
"Zhigang",
""
],
[
"Long",
"Bo",
""
],
[
"Zhao",
"Tong",
""
],
[
"Shah",
"Neil",
""
],
[
"Javari",
"Amin",
""
],
[
"Xia",
"Yinglong",
""
],
[
"Tang",
"Jiliang",
""
]
]
}
|
## RAG vs. GraphRAG: A Systematic Evaluation and Key Insights
### Haoyu Han[1], Yu Wang[2], Harry Shomer[1], Yongjia Lei, [2], Kai Guo [1], Zhigang Hua[3], Bo Long [3], Hui Liu[1], Jiliang Tang[1]
1Michigan State University, 2University of Oregon, 3Meta
### {hanhaoy1, shomerha, guokai1, liuhui7, tangjili}@msu.edu {yuwang, yongjia}@uoregon.edu, {zhua, bolong}@meta.com
### Abstract
Retrieval-Augmented Generation (RAG) enhances the performance of LLMs across various tasks by retrieving relevant information
from external sources, particularly on textbased data. For structured data, such as knowledge graphs, GraphRAG has been widely used
to retrieve relevant information. However, recent studies have revealed that structuring implicit knowledge from text into graphs can
benefit certain tasks, extending the application of GraphRAG from graph data to general
text-based data. Despite their successful extensions, most applications of GraphRAG for
text data have been designed for specific tasks
and datasets, lacking a systematic evaluation
and comparison between RAG and GraphRAG
on widely used text-based benchmarks. In
this paper, we systematically evaluate RAG
and GraphRAG on well-established benchmark
tasks, such as Question Answering and Querybased Summarization. Our results highlight
the distinct strengths of RAG and GraphRAG
across different tasks and evaluation perspectives. Inspired by these observations, we investigate strategies to integrate their strengths
to improve downstream tasks. Additionally,
we provide an in-depth discussion of the shortcomings of current GraphRAG approaches and
outline directions for future research.
### 1 Introduction
Retrieval-Augmented Generation (RAG) has
emerged as a powerful approach to enhance downstream tasks by retrieving relevant knowledge from
external data sources. It has achieved remarkable
success in various real-world applications, such
as healthcare (Xu et al., 2024), law (Wiratunga
et al., 2024), finance (Zhang et al., 2023), and education (Miladi et al., 2024). This success has been
further amplified with the advent of Large Language Models (LLMs), as integrating RAG with
LLMs significantly improves their faithfulness by
mitigating hallucinations, reducing privacy risks,
and enhancing robustness (Zhao et al., 2023; Huang
et al., 2023). In most existing RAG systems, retrieval is primarily conducted from text databases
using lexical and semantic search.
Graphs, as a fundamental data structure, encode
rich relational information and have been extensively utilized across real-world domains, including
knowledge representation, social network analysis,
and biomedical research (Wu et al., 2020; Ma and
Tang, 2021; Wu et al., 2023). Motivated by this,
GraphRAG has recently gained attention for retrieving graph-structured data, such as knowledge
graphs (KGs) and molecular graphs (Han et al.,
2024; Peng et al., 2024). Beyond leveraging existing graphs, GraphRAG has also demonstrated its
effectiveness for text-based tasks after structuring
implicit knowledge from text into graph representations, benefiting applications such as global summarization (Edge et al., 2024; Zhang et al., 2024),
planning (Lin et al., 2024) and reasoning (Han et al.,
2025).
While previous studies have demonstrated the
potential of GraphRAG for text-based tasks by
converting sequential text into graphs, most of
them primarily focus on specific tasks and welldesigned datasets. Consequently, the applicability
of GraphRAG to broader, real-world text-based
tasks remains unclear, particularly when compared
to RAG, which has seen widespread adoption
across diverse applications. This raises a critical
question: What are the advantages and disadvan_tages of applying GraphRAG to general text-based_
_tasks compared to RAG?_
To bridge this gap, we systematically evaluate
the performance of RAG and GraphRAG on general text-based tasks using widely adopted datasets,
including Question Answering and Query-based
Summarization. Specifically, we assess two representative GraphRAG methods: (1) Knowledge
Graph-based GraphRAG (Liu, 2022), which ex
-----
tracts a Knowledge Graph (KG) from text and performs retrieval solely based on the KG and (2)
Community-based GraphRAG (Edge et al., 2024),
which retrieves information not only from the constructed KG but also from hierarchical communities within the graph. For the Question Answering task, we conduct experiments on both singlehop and multi-hop QA under single-document and
multi-document scenarios. Similarly, for the Querybased Summarization task, we evaluate both singledocument and multi-document summarization to
comprehensively assess the effectiveness of RAG
and GraphRAG.
Based on our comprehensive evaluation, we
conduct an in-depth analysis of the strengths and
weaknesses of RAG and GraphRAG across different tasks. Our findings reveal that RAG and
GraphRAG are complementary, each excelling in
different aspects. For the Question Answering task,
we observe that RAG performs better on singlehop questions and those requiring detailed information, while GraphRAG is more effective for
multi-hop questions. In the Query-based Summarization task, RAG captures fine-grained details,
whereas GraphRAG generates more diverse and
multi-faceted summaries. Building on these insights, we investigate two strategies from different
perspectives to integrate their unique strengths and
enhance the overall performance. Our main contributions are as follows:
- Systematical Evaluation : This is the very first
work to systematically evaluate and compare
RAG and GraphRAG on text-based tasks using
widely adopted datasets and evaluations.
- Task-Specific Insights: We provide an in-depth
analysis of the distinct strengths of RAG and
GraphRAG, demonstrating their complementary
advantages across different types of queries and
objectives.
- Hybrid Retrieval Strategies: Based on our
findings on the unique strengths of RAG and
GraphRAG, we propose two strategies to improve overall performance: (1) Selection, where
queries are dynamically assigned to either RAG
or GraphRAG based on their characteristics, and
(2) Integration, where both methods are integrated to leverage their complementary strengths.
- Challenges and Future Directions: We discuss
the limitations of current GraphRAG approaches
and outline potential future research directions
for broader applicability.
### 2 Related Works
**2.1** **Retrieval-Augmented Generation**
Retrieval-Augmented Generation (RAG) has been
widely applied to enhance the performance of
Large Language Models (LLMs) by retrieving relevant information from external sources, addressing
the limitation of LLMs’ restricted context windows,
improving factual accuracy, and mitigating hallucinations (Fan et al., 2024; Gao et al., 2023). Most
RAG systems primarily process text data by first
splitting it into chunks (Finardi et al., 2024). When
a query is received, RAG retrieves relevant chunks
either through lexical search (Ram et al., 2023)
or by computing semantic similarity (Karpukhin
et al., 2020), embeddings both the query and text
chunks into a shared vector space. Advanced techniques, such as pre-retrieval processing (Ma et al.,
2023; Zheng et al., 2023a) and post-retrieval processing (Dong et al., 2024; Xu et al., 2023), as
well as fine-tuning strategies (Li et al., 2023), have
further enhanced RAG’s effectiveness across various domains, including QA) (Yan et al., 2024),
dialogue generation (Izacard et al., 2023), and text
summarization (Jiang et al., 2023).
Several studies have evaluated the effectiveness
of RAG systems across various tasks (Yu et al.,
2024; Chen et al., 2024; Es et al., 2023), such
as multi-hop question answering (Tang and Yang,
2024), biomedical question answering (Xiong et al.,
2024), and text generation (Liu et al., 2023). However, no existing study has simultaneously and
systematically evaluated and compared RAG and
GraphRAG on these general text-based tasks.
**2.2** **Graph Retrieval-Augmented Generation**
While RAG primarily processes text data, many
real-world scenarios involve graph-structured data,
such as knowledge graphs (KGs), social graphs,
and molecular graphs (Xia et al., 2021; Ma and
Tang, 2021). GraphRAG (Han et al., 2024; Peng
et al., 2024) aims to retrieve information from various types of graph-structured data. The inherent
structure of graphs enhances retrieval by capturing relationships between connected nodes. For
example, hyperlinks between documents can improve retrieval effectiveness in question answering
tasks(Li et al., 2022). Currently, most GraphRAG
studies focus on retrieving information from existing KGs for downstream tasks such as KG-based
QA (Tian et al., 2024; Yasunaga et al., 2021) and
Fact-Checking (Kim et al., 2023).
-----
Figure 1: The illustration of RAG, KG-based GraphRAGs and Community-based GraphRAGs.
Despite leveraging the existing graphs, recent
studies have explored incorporating graph construction into GraphRAG to enhance text-based
tasks. For example, Dong et al. (2024) construct
document graphs using Abstract Meaning Representation (AMR) to improve document ranking.
Edge et al. (2024) construct graphs from documents
using LLMs, where nodes represent entities and
edges capture relationships between them. Based
on these graphs, they generate hierarchical communities and corresponding community summaries
or reports. Their approach focuses on the global
query summarization task, retrieving information
from both the constructed graphs and their hierarchical communities. Additionally, Han et al. (2025)
propose an iterative graph construction approach
using LLMs to improve reasoning tasks.
These studies highlight the potential of
GraphRAG in processing text-based tasks by constructing graphs from textual data. However, their
focus is limited to specific tasks and evaluation
settings. It remains unclear how GraphRAG performs on general text-based tasks compared to
RAG. More importantly, when and how should
GraphRAG be applied to such tasks for optimal
effectiveness? Our work aims to bridge this gap by
systematically evaluating GraphRAG and comparing it with RAG on general text-based tasks.
### 3 Evaluation Methodology
In this section, we introduce the details of our
evaluation framework. We primarily evaluate one
representative RAG system and two representative
GraphRAG systems, as illustrated in Figure 1.
**3.1** **RAG**
We adopt a representative semantic similaritybased retrieval approach as our RAG
method (Karpukhin et al., 2020). Specifically, we
first split the text into chunks, each containing
approximately 256 tokens. For indexing, we use
OpenAI’s text-embedding-ada-002 model, which
has demonstrated effectiveness across various
tasks (Nussbaum et al., 2024). For each query, we
retrieve chunks with Top-10 similarity scores. To
generate responses, we employ two open-source
models of different sizes: Llama-3.1-8B-Instruct
and Llama-3.1-70B-Instruct (Dubey et al., 2024).
For single-document tasks, we generate a separate RAG system for each document, ensuring that
queries corresponding to a specific document are
processed within its respective indexed chunk pool.
For multi-document tasks, we use a shared RAG
system by indexing all documents together.
**3.2** **GraphRAG**
We select two representative GraphRAG methods for a comprehensive evaluation, as shown
in Figure 1, namely KG-based GraphRAG and
Community-based GraphRAG.
In the KG-based GraphRAG (KGGraphRAG) (Liu, 2022), a knowledge graph is first
constructed from text chunks using LLMs through
triplet extraction. When a query is received, its
entities are extracted and matched to those in
the constructed KG using LLMs. The retrieval
process then traverses the graph from the matched
entities and gathers triplets (head, relation, tail)
from their multi-hop neighbors as the retrieved
-----
content. Additionally, for each triplet, we can
retrieve the corresponding text associated with
it. We define two variants of KG-GraphRAG: (1)
_KG-GraphRAG (Triplets), which retrieves only the_
triplets, and (2) KG-GraphRAG (Triplets+Text),
which retrieves both the triplets and their associated
source text. We implement the KG-GraphRAG
methods using LlamaIndex (Liu, 2022) [1].
For the Community-based GraphRAG (Edge
et al., 2024), in addition to generating KGs using
LLMs, hierarchical communities are constructed
using graph community detection algorithms, as
shown in Figure 1. Each community is associated with a corresponding text summary or report,
where lower-level communities contain detailed
information from the original text. The higherlevel communities further provide summaries of
the lower-level communities. Due to the hierarchical community structure, there are two primary
retrieval methods for retrieving relevant information given a query: Local Search and Global
**Search. In Local Search, entities, relations, their**
descriptions, and lower-level community reports
are retrieved based on entity matching between the
query’s extracted entities and the constructed graph.
We refer to this method as Community-GraphRAG
_(Local). In Global Search, only high-level com-_
munity summaries are retrieved based on semantic
similarity to the query. We refer to this method as
_Community-GraphRAG (Global). The Community-_
GraphRAG methods are implemented using Microsoft GraphRAG (Edge et al., 2024)[2].
To ensure a fair comparison, we adopt the same
settings for both RAG and GraphRAG methods.
This includes the chunking strategy, embedding
model, and LLMs. We select two representative RAG tasks, i.e., Question Answering and
Query-based Summarization, to evaluate RAG and
GraphRAG simultaneously.
### 4 Question Answering
QA is one of the most widely used tasks for evaluating the performance of RAG systems. QA tasks
come in various forms, such as single-hop QA,
multi-hop QA, and open-domain QA (Wang, 2022).
To systematically assess the effectiveness of RAG
and GraphRAG in these tasks, we evaluate them
on widely used QA datasets and employ standard
evaluation metrics.
1https://www.llamaindex.ai/
2https://microsoft.github.io/graphrag
**4.1** **Datasets and Evaluation Metrics**
To comprehensively evaluate the performance of
GraphRAG on general QA tasks, we select four
widely used datasets that cover different perspectives. For the single-hop QA task, we select
the Natural Questions (NQ) dataset (Kwiatkowski
et al., 2019). For the multi-hop QA task, we select HotPotQA (Yang et al., 2018) and MultiHopRAG (Tang and Yang, 2024) datasets. The
MultiHop-RAG dataset categorizes queries into
four types: Inference, Comparison, Temporal, and
Null queries. To further analyze the performance of
RAG and GraphRAG at a finer granularity, we also
include NovelQA (Tang and Yang, 2024), which
contains 21 different types of queries. For more
details, please refer to Appendix A.1.1. We use
Precision (P), Recall (R), and F1-score as evaluation metrics for the NQ and HotPotQA datasets,
while accuracy is used for the MultiHop-RAG and
NovelQA datasets following their original papers.
**4.2** **QA Main Results**
The performance comparison for the NQ and HotPotQA datasets is presented in Table 1, while that
of MultiHop-RAG is shown in Table 2. Due to
space constraints, partial results of NovelQA with
the Llama 3.1-8B model are shown in Table 3, with
the full results available in Appendix A.2. Based on
these results, we make the following observations:
1. RAG excels on detailed single-hop queries.
RAG performs well on single-hop queries and
queries that require detailed information. This
is evident from its performance on the singlehop dataset (NQ) as well as the single-hop (sh)
and detail-oriented (dtl) queries in the NovelQA
dataset, as shown in Table 1 and Table 3.
2. GraphRAG, **particularly** **Community-**
**GraphRAG (Local), excels on multi-hop**
**queries.** For instance, it achieved the best
performance on both the HotPotQA and
MultiHop-RAG datasets. Although its overall
performance on the NovelQA dataset is lower
than that of RAG, it still performs well on the
multi-hop (mh) queries in NovelQA dataset.
3. Community-GraphRAG (Global) often strug**gles on QA tasks. This is due to the global**
search retrieves only high-level communities,
leading to a loss of detailed information. This is
particularly evident from its lower performance
on detail-oriented queries in the NovelQA
dataset. Additionally, Community-GraphRAG
(Global) tends to hallucinate in QA tasks, as
-----
Table 1: Performance comparison (%) on NQ and Hotpot datasets. The best results are highlighted in bold, and the
second-best results are underlined.
**NQ** **Hotpot**
**Method** **Llama 3.1-8B** **Llama 3.1-70B** **Llama 3.1-8B** **Llama 3.1-70B**
P R F1 P R F1 P R F1 P R F1
RAG **71.7** **63.93** **64.78** **74.55** **67.82** **68.18** 62.32 60.47 60.04 66.34 63.99 63.88
KG-GraphRAG (Triplets only) 40.09 33.56 34.28 37.84 31.22 28.50 26.88 24.81 25.02 32.59 30.63 30.73
KG-GraphRAG (Triplets+Text) 58.36 48.93 50.27 60.91 52.75 53.88 45.22 42.85 42.60 51.44 48.99 48.75
Community-GraphRAG (Local) 69.48 62.54 63.01 71.27 65.46 65.44 **64.14** **62.08** **61.66** **67.20** **64.89** **64.60**
Community-GraphRAG (Global) 60.76 54.99 54.48 61.15 55.52 55.05 45.72 47.60 45.16 48.33 48.56 46.99
Table 2: Performance comparison (%) on the MultiHop-RAG dataset across different query types.
**LLama 3.1-8B** **Llama 3.1-70B**
**Method**
**Inference** **Comparison** **Null** **Temporal** **Overall** **Inference** **Comparison** **Null** **Temporal** **Overall**
RAG **92.16** 57.59 96.01 30.7 67.02 **94.85** 56.31 91.36 25.73 65.77
KG-GraphRAG (Triplets only) 55.76 22.55 **98.67** 18.7 41.24 76.96 32.36 **94.35** 19.55 50.98
KG-GraphRAG (Triplets+Text) 67.4 34.7 97.34 17.15 48.51 85.91 35.98 86.38 21.61 54.58
Community-GraphRAG (Local) 86.89 60.63 80.07 50.6 **69.01** 92.03 60.16 88.70 49.06 **71.17**
Community-GraphRAG (Global) 89.34 **64.02** 19.27 **53.34** 64.4 89.09 **66.00** 13.95 **59.18** 65.69
Table 3: Performance comparison (%) on the NovelQA dataset across different query types with LLama 3.1-8B.
|NQ Hotpot Method Llama 3.1-8B Llama 3.1-70B Llama 3.1-8B Llama 3.1-70B P R F1 P R F1 P R F1 P R F1|NQ|Col3|Hotpot|Col5|
|---|---|---|---|---|
||Llama 3.1-8B|Llama 3.1-70B|Llama 3.1-8B|Llama 3.1-70B|
||P R F1|P R F1|P R F1|P R F1|
|RAG KG-GraphRAG (Triplets only) KG-GraphRAG (Triplets+Text) Community-GraphRAG (Local) Community-GraphRAG (Global)|71.7 63.93 64.78 40.09 33.56 34.28 58.36 48.93 50.27 69.48 62.54 63.01 60.76 54.99 54.48|74.55 67.82 68.18 37.84 31.22 28.50 60.91 52.75 53.88 71.27 65.46 65.44 61.15 55.52 55.05|62.32 60.47 60.04 26.88 24.81 25.02 45.22 42.85 42.60 64.14 62.08 61.66 45.72 47.60 45.16|66.34 63.99 63.88 32.59 30.63 30.73 51.44 48.99 48.75 67.20 64.89 64.60 48.33 48.56 46.99|
|LLama 3.1-8B Llama 3.1-70B Method Inference Comparison Null Temporal Overall Inference Comparison Null Temporal Overall|LLama 3.1-8B|Llama 3.1-70B|
|---|---|---|
||Inference Comparison Null Temporal Overall|Inference Comparison Null Temporal Overall|
|RAG KG-GraphRAG (Triplets only) KG-GraphRAG (Triplets+Text) Community-GraphRAG (Local) Community-GraphRAG (Global)|92.16 57.59 96.01 30.7 67.02 55.76 22.55 98.67 18.7 41.24 67.4 34.7 97.34 17.15 48.51 86.89 60.63 80.07 50.6 69.01 89.34 64.02 19.27 53.34 64.4|94.85 56.31 91.36 25.73 65.77 76.96 32.36 94.35 19.55 50.98 85.91 35.98 86.38 21.61 54.58 92.03 60.16 88.70 49.06 71.17 89.09 66.00 13.95 59.18 65.69|
|RAG|KG-GraphRAG (Triplets+Text)|
|---|---|
|chara mean plot relat settg span times avg mh 68.75 52.94 58.33 75.28 92.31 64.00 33.96 47.34 sh 69.08 62.86 66.11 75.00 78.35 - - 68.73 dtl 64.29 45.51 78.57 10.71 83.78 - - 55.28 avg 67.78 50.57 67.37 60.80 80.95 64.00 33.96 57.12|chara mean plot relat settg span times avg mh 52.08 52.94 44.44 55.06 69.23 64.00 28.61 38.37 sh 36.84 45.71 40.17 87.50 36.08 - - 39.93 dtl 38.57 30.90 42.86 21.43 32.43 - - 33.60 avg 40.00 36.23 41.09 49.60 38.10 64.00 28.61 37.80|
|Community-GraphRAG (Local)|Community-GraphRAG (Global)|
|chara mean plot relat settg span times avg mh 68.75 64.71 55.56 67.42 92.31 52.00 35.83 47.01 sh 59.87 58.57 65.69 87.50 64.95 - - 63.43 dtl 54.29 37.64 62.50 25.00 70.27 - - 46.88 avg 60.00 44.91 64.05 59.20 68.71 52.00 35.83 53.03|chara mean plot relat settg span times avg mh 54.17 58.82 55.56 56.18 53.85 68.00 20.59 34.39 sh 45.39 50.00 55.65 87.50 38.14 - - 49.65 dtl 28.57 29.78 32.14 87.50 40.54 - - 30.89 avg 42.59 36.98 51.66 52.00 40.14 68.00 20.59 39.17|
shown by its poor performance on Null queries
in the MultiHop-RAG dataset, which should ideally be answered as ‘insufficient information.’
However, this summarization approach may be
beneficial for queries that require comparing
different topics or understanding their temporal ordering, such as Comparison and Temporal
queries in the MultiHop-RAG dataset, as shown
in Table 2.
4. KG-based GraphRAG also generally under**perform on QA tasks. This is because it re-**
trieves information solely from the constructed
knowledge graph, which contains only entities
and their relations. However, the extracted entities and relations may be incomplete, leading
to gaps in the retrieved information. To verify
this, we calculated the ratio of answer entities
present in the constructed KG. We found that
only around 65.8% of answer entities exist in
the constructed KG for the Hotpot dataset and
65.5% for the NQ dataset. These findings highlight a key limitation in KG-based retrieval and
suggest the need for improved KG construction
methods to enhance graph completeness for QA.
**4.3** **Comparative QA Analysis**
In this section, we conduct a detailed analysis of
the behavior of RAG and GraphRAG, focusing
on their strengths and weaknesses. In the following discussion, we refer to Community-GraphRAG
(Local) as GraphRAG, as it demonstrates performance comparable to RAG. We categorize queries
into four groups: (1) Queries correctly answered
by both methods, (2) Queries correctly answered
only by RAG (RAG-only), (3) Queries correctly answered only by GraphRAG (GraphRAG-only), and
**(4) Queries answered incorrectly by both methods.**
The confusion matrices representing these four
groups using the Llama 3.1-8B model are shown
in Figure 2. Notably, the proportions of queries
correctly answered exclusively by GraphRAG and
RAG are significant. For example, 13.6% of
queries are GraphRAG-only, while 11.6% are RAGonly on MultiHop-RAG dataset. This phenomenon
highlights the complementary properties of RAG
and GraphRAG, and each method has its own
strengths and weaknesses. Therefore, leveraging
_their unique advantages has the potential to im-_
_prove overall performance._
-----
40.0 17.1
13.7 29.1
45.4 9.2
9.8 35.6
(a) NQ
47.2 7.8
6.0 39.0
(b) Hotpot
55.4 11.6
13.6 19.4
(c) MultiHop-RAG
(d) NovelQA
Figure 2: Confusion matrices comparing GraphRAG and RAG correctness across datasets using Llama 3.1-8B.
**4.4** **Improving QA Performance**
Building on the complementary properties of RAG
and GraphRAG, we investigate the following two
strategies to enhance overall QA performance.
**Strategy 1: RAG vs. GraphRAG Selection.**
In Section 4.2, we observe that RAG generally
performs well on single-hop queries and those
requiring detailed information, while GraphRAG
(Community-GraphRAG (Local)) excels in multihop queries that require reasoning. Therefore, we
hypothesize that RAG is well-suited for fact-based
queries, which rely on direct retrieval and detailed
information, whereas GraphRAG is more effective
for reasoning-based queries that involve chaining
multiple facts together. Therefore, given a query,
we employ a classification mechanism to determine
whether it is fact-based or reasoning-based. Each
query is then assigned to either RAG or GraphRAG
based on the classification results. Specifically, we
leverage the in-context learning ability of LLMs
for classification (Dong et al., 2022; Wei et al.,
2023). Further details and prompts can be found
in Appendix A.3. In this strategy, either RAG or
GraphRAG is selected for each query, and we refer
to this strategy as Selection.
**Strategy 2: RAG and GraphRAG Integration.**
We also explore the Integration strategy to leverage the complementary strengths of RAG and
GraphRAG. Both RAG and GraphRAG retrieve
information for a query simultaneously. The retrieved results are then concatenated and fed into
the generator to produce the final output.
We conduct experiments to verify the effectiveness of the two proposed strategies. Specifically,
we evaluate overall performance across all selected
datasets. For the MultiHop-RAG and NovelQA
datasets, we use the overall accuracy, while for the
NQ and HotPotQA datasets, we use the F1 score
as the evaluation metric. The results are shown
in Figure 3. From these results, we observe that
**both strategies generally enhance overall per-**
**formance. For example, on the MultiHop-RAG**
dataset with Llama 3.1-70B, Selection and Integration improve the best method by 1.1% and 6.4%,
respectively. When comparing the Selection and
Integration strategies, the Integration strategy usually achieves higher performance than the Selection strategy. However, the Selection strategy processes each query using either RAG or GraphRAG,
making it more efficient. In contrast, the Integration strategy yields better performance but requires each query to be processed by both RAG
and GraphRAG, increasing computational cost.
### 5 Query-Based Summarization
Query-based summarization tasks are widely used
to evaluate the performance of RAG systems (Ram
et al., 2023; Yu et al., 2023). GraphRAG has
also demonstrated its effectiveness in summarization tasks (Edge et al., 2024). However, Edge
et al. (2024) only evaluate its effectiveness on the
global summarization task and rely on LLM-as-aJudge (Zheng et al., 2023b) for performance assessment. In Section 5.3, we show that the LLMas-a-Judge evaluation method for summarization
tasks introduces position bias, which can impact
the reliability of results. A systematic comparison
of RAG and GraphRAG on general query-based
summarization across widely used datasets remains
unexplored. To address this gap, we conduct a comprehensive evaluation in this section, leveraging
widely used datasets and evaluation metrics.
**5.1** **Datasets and Evaluation Metrics**
We adopt two widely used single-document querybased summarization datasets, SQuALITY (Wang
et al., 2022) and QMSum (Zhong et al., 2021),
and two multi-document query-based summarization datasets, ODSum-story and ODSummeeting (Zhou et al., 2023), for our evaluation.
Unlike the LLM-generated global queries used in
-----
70
65
75
70
60
55
65
60
50
NQ Hotpot MultiHop-RAG NovelQA
55
50
NQ Hotpot MultiHop-RAG NovelQA
(a) Llama3.1-8B (b) Llama3.1-70B
Figure 3: Overall QA performance comparison of different methods.
the unreleased datasets of Edge et al. (2024), most
queries in the selected datasets focus on specific
roles or events. Since these datasets contain one or
more ground truth summaries for each query, we
use ROUGE-2 (Lin, 2004) and BERTScore (Zhang
et al., 2019) as evaluation metrics to measure lexical and semantic similarity between the predicted
and ground truth summaries.
**5.2** **Summarization Experimental Results**
We evaluate both the KG-based and Communitybased GraphRAG methods, along with the Integration strategy discussed in Section 4.4. The results of Llama3.1-8B model on Query-based single
document summarization and multiple document
summarization are shown in Table 4 and Table 5, respectively. The results of Llama3.1-70B are shown
in Appendix A.4. Based on these results, we can
make the following observations:
1. RAG generally performs well on query-based
**summarization tasks.** This is particularly
true on multi-document summarization datasets,
where they are often the best method.
2. KG-based GraphRAG benefit from combin**ing triplets with their corresponding text.**
This improves performance by incorporating
more details, making predictions closer to the
ground truth summaries.
3. Community-based GraphRAG performs bet**ter with the Local search method.** Local
search retrieves entities, relations, and lowlevel communities, while the Global search
method retrieves only high-level summaries.
This demonstrates the importance of detailed
information in the selected datasets.
4. The Integration strategy is often comparable
**to RAG only performance. This strategy in-**
tegrates retrieved content from both RAG and
Community-GraphRAG (Local), resulting in
performance similar to RAG alone.
**5.3** **Position Bias in Existing Evaluation**
From the results in Section 5.2, the Communitybased GraphRAG, particularly with global search,
generally underperforms compared to RAG on the
selected datasets. This contrasts with the findings
of Edge et al. (2024), where Community-based
GraphRAG with global search outperformed both
local search and RAG. There are two key differences between our evaluation and Edge et al.
(2024). First, their study primarily focuses on
global summarization, which captures the overall
information of an entire corpus, whereas the selected datasets in our evaluation contain queries related to specific roles or events. Second, Edge et al.
(2024) assess performance by comparing RAG
and GraphRAG outputs using LLM-as-a-Judge
without ground truth, whereas we evaluate results
against ground truth summaries using ROUGE and
BERTScore. These metrics emphasize similarity
to the reference summaries, which often contain
more detailed information.
We further conduct an evaluation following Edge
et al. (2024), using the LLM-as-a-Judge method to
compare RAG and Community-based GraphRAG
from two perspectives: Comprehensiveness and
Diversity. Comprehensiveness focuses on detail,
addressing the question: "How much detail does
_the answer provide to cover all aspects and details_
_of the question?" Meanwhile, Diversity emphasizes_
global information, evaluating "Does the answer
_provide a broad and globally inclusive perspec-_
_tive?". The prompt and details are shown in Ap-_
pendix A.5. Specifically, we input the summaries
generated by RAG and GraphRAG into the prompt
and ask the LLM to select the better one for each
metric, following Edge et al. (2024). Additionally,
to better account for the order in which the summaries are presented, we consider two scenarios.
_Order 1 (O1): We place the RAG summary appears_
|RAG GraphRAG|Col2|Col3|Col4|Col5|Col6|Col7|Col8|Col9|Col10|Col11|Col12|Col13|Col14|Col15|Col16|Col17|Col18|Col19|Col20|Col21|Col22|
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
|||||||||||||||||||||||
|Selection Integration||||||||||||||||||||||
|||||||||||||||||||||||
|||||||||||||||||||||||
|||||||||||||||||||||||
|||||||||||||||||||||||
|||||||||||||||||||||||
|||||||||||||||||||||||
|||||||||||||||||||||||
|a m e e c E p al n 2 b|ri P t d ra te dg h s c 4) al|N son os he Gr lly d e RA ea es . su|Q o iti r ap u dat et G rc be Fi m|f d on es hR nd as al w h tw rst ma|iff B ult A er et . ( ith an e , ri||(b er ia s i G, pe s. 2 g d en th za|Hot ) L ent s i n p rfo T 02 lo R o eir tio|po la m n Se ar rm his 4) ba A ur s n,|t ma et E ct tic s c , l s G. e tu w|M 3.1 ho xis io ul co on wh ea T val dy hi|ult -7 ds. ti n arl m tra er rc he ua p ch|iH 0B ng 5. y pa st e h re ti ri c|op- E 2, wi re s Co ou a on m ap|RA va th th d wi m tp re a ari tu|G lu e gl to th m er t nd ly re|N at Co o R th un for w E f s t|ov io m bal AG e it m o d oc he|elQ n m s o fin y- ed ke ge us o|A un ea n di ba b y et es ve|it rc th ng se ot di a o ra|
|RAG GraphRAG|Col2|Col3|Col4|Col5|Col6|Col7|Col8|Col9|Col10|Col11|Col12|Col13|Col14|Col15|Col16|Col17|Col18|Col19|Col20|Col21|Col22|
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
|Selection Integration||||||||||||||||||||||
|||||||||||||||||||||||
|||||||||||||||||||||||
|||||||||||||||||||||||
|||||||||||||||||||||||
|||||||||||||||||||||||
|||||||||||||||||||||||
|||||||||||||||||||||||
|||||||||||||||||||||||
|n e o O , n r S v o o m|re s r gr U 20 d s ou u al G n f e|N lea in eve ou G 19 e nd m uat ra str Lla nt|Q se th nt nd E- ) ma tr m e ph ate m su|d e s s. tr 2 ( as nt ut ari bo R gy a3 m|da el Si ut Li ev ic h za th AG d .1- ma||(a tas ec nc h n, al si su ti th is 8 ri|Hot ) F et te e su 2 ua mi m on e me cu B za|po Lla ig s o d th m 00 tio lar ma E K th ss mo tio|t ma ure f da ese ma 4) n it ri x G- od ed de n|M 3. 3 Ed tas d ri an m y b es. pe ba s, in l an|ult 1-8 : O ge et at es d etr et ri se a S on d|iH B ve e s as fo B ic w me d lo ec Q m|op- ra t a fo ets r ER s t ee nt an ng ti ue ul|RA ll l. cu c ea T o n t al d w on ry tip|G QA (2 s on ch Sc me he R C it 4 -b le|N p 02 on tai q or as p es om h .4. as d|ov er 4) sp n ue e ( ur re ul m th T ed oc|elQ for , e on ry, Zh e di ts u e I h si um|A m mo cif e w an lex cte nit nt e r ng e|an st ic or e g i- d y- e- e- le nt|
-----
Table 4: The performance of query-based single document summarization task using Llama3.1-8B.
**SQuALITY** **QMSum**
**ROUGE-2** **BERTScore** **ROUGE-2** **BERTScore**
P R F1 P R F1 P R F1 P R
15.09 8.74 10.08 74.54 81.00 77.62 21.50 **3.80** 6.32 **81.03** 84.45
KG-GraphRAG (Triplets only) 11.99 6.16 7.41 82.46 84.30 83.17 13.71 2.55 4.15 80.16 82.96
KG-GraphRAG (Triplets+Text) 15.00 **9.48** 10.52 **84.37** **85.88** **84.92** 16.83 3.32 5.38 80.92 83.64
Community-GraphRAG (Local) **15.82** 8.64 10.10 83.93 85.84 84.66 20.54 3.35 5.64 80.63 84.13
Community-GraphRAG (Global) 10.23 6.21 6.99 82.68 84.26 83.30 10.54 1.97 3.23 79.79 82.47
Integration 15.69 9.32 **10.67** 74.56 81.22 77.73 **21.97** **3.80** **6.34** 80.89 **84.47**
Table 5: The performance of query-based multiple document summarization task using Llama3.1-8B.
|SQuALITY QMSum Method ROUGE-2 BERTScore ROUGE-2 BERTScore P R F1 P R F1 P R F1 P R F1|SQuALITY|QMSum|
|---|---|---|
||ROUGE-2 BERTScore|ROUGE-2 BERTScore|
||P R F1 P R F1|P R F1 P R F1|
|RAG KG-GraphRAG (Triplets only) KG-GraphRAG (Triplets+Text) Community-GraphRAG (Local) Community-GraphRAG (Global) Integration|15.09 8.74 10.08 74.54 81.00 77.62 11.99 6.16 7.41 82.46 84.30 83.17 15.00 9.48 10.52 84.37 85.88 84.92 15.82 8.64 10.10 83.93 85.84 84.66 10.23 6.21 6.99 82.68 84.26 83.30 15.69 9.32 10.67 74.56 81.22 77.73|21.50 3.80 6.32 81.03 84.45 82.69 13.71 2.55 4.15 80.16 82.96 81.52 16.83 3.32 5.38 80.92 83.64 82.25 20.54 3.35 5.64 80.63 84.13 82.34 10.54 1.97 3.23 79.79 82.47 81.10 21.97 3.80 6.34 80.89 84.47 82.63|
|ODSum-story ODSum-meeting Method ROUGE-2 BERTScore ROUGE-2 BERTScore P R F1 P R F1 P R F1 P R F1 RAG 15.39 8.44 9.81 83.87 85.74 84.57 15.50 6.43 8.77 83.12 85.84 84.45 KG-GraphRAG (Triplets only) 11.02 5.56 6.62 82.09 83.91 82.77 11.64 4.87 6.58 81.13 84.32 82.69 KG-GraphRAG (Triplets+Text) 9.19 5.82 6.22 79.39 83.30 81.03 11.97 4.97 6.72 81.50 84.41 82.92 Community-GraphRAG (Local) 13.84 7.19 8.49 83.19 85.07 83.90 15.65 5.66 8.02 82.44 85.54 83.96 Community-GraphRAG (Global) 9.40 4.47 5.46 81.46 83.54 82.30 11.44 3.89 5.59 81.20 84.50 82.81 Integration 14.77 8.55 9.53 83.73 85.56 84.40 15.69 6.15 8.51 82.87 85.81 84.31|ODSum-story|ODSum-meeting|
|---|---|---|
||ROUGE-2 BERTScore|ROUGE-2 BERTScore|
||P R F1 P R F1|P R F1 P R F1|
||15.39 8.44 9.81 83.87 85.74 84.57 11.02 5.56 6.62 82.09 83.91 82.77 9.19 5.82 6.22 79.39 83.30 81.03 13.84 7.19 8.49 83.19 85.07 83.90 9.40 4.47 5.46 81.46 83.54 82.30 14.77 8.55 9.53 83.73 85.56 84.40|15.50 6.43 8.77 83.12 85.84 84.45 11.64 4.87 6.58 81.13 84.32 82.69 11.97 4.97 6.72 81.50 84.41 82.92 15.65 5.66 8.02 82.44 85.54 83.96 11.44 3.89 5.59 81.20 84.50 82.81 15.69 6.15 8.51 82.87 85.81 84.31|
1.0
0.8
1.0
0.8
1.0
0.8
1.0
0.8
0.6
0.4
0.6
0.4
0.6
0.4
0.6
0.4
0.2
0.0
Comprehensiveness Diversity
0.2
0.0
Comprehensiveness Diversity
0.2
0.0
Comprehensiveness Diversity
0.2
0.0
Comprehensiveness Diversity
|RAG-Order 1 GraphRAG-Local-Order 1 RAG-Order 2 GraphRAG-Local-Order 2|Col2|Col3|Col4|RAG-Order 1 GraphRAG-Local-Order 1 RAG-Order 2 GraphRAG-Local-Order 2|Col6|Col7|Col8|Col9|Col10|Col11|
|---|---|---|---|---|---|---|---|---|---|---|
||||||G R G||||||
||||||||||||
||||||||||||
||||||||||||
||||||||||||
||||||||||||
||||||||||||
||||||||||||
|Col1|Col2|RAG-Order 1 GraphRAG-Gloabl-Order 1 RAG-Order 2 GraphRAG-Gloabl-Order 2|RAG-Order 1 GraphRAG-Gloabl-Order 1 RAG-Order 2 GraphRAG-Gloabl-Order 2|Col5|Col6|Col7|Col8|
|---|---|---|---|---|---|---|---|
|||||||||
|||||||||
|RAG-Order 1 GraphRAG-Local-Order 1 RAG-Order 2 GraphRAG-Local-Order 2|Col2|Col3|Col4|RAG-Order 1 GraphRAG-Local-Order 1 RAG-Order 2 GraphRAG-Local-Order 2|Col6|Col7|Col8|Col9|Col10|Col11|
|---|---|---|---|---|---|---|---|---|---|---|
||||||||||||
||||||||||||
||||||||||||
||||||||||||
||||||||||||
||||||||||||
|Col1|Col2|RAG-Order 1 GraphRAG-Gloabl-Order 1 RAG-Order 2 GraphRAG-Gloabl-Order 2|RAG-Order 1 GraphRAG-Gloabl-Order 1 RAG-Order 2 GraphRAG-Gloabl-Order 2|
|---|---|---|---|
(a) QMSum Local (b) QMSum Global (c) ODSum-story Local (d) ODSum-story Global
Figure 4: Comparison of LLM-as-a-Judge evaluations for RAG and GraphRAG. "Local" refers to the evaluation of
RAG vs. GraphRAG-Local, while "Global" refers to RAG vs. GraphRAG-Global.
before the GraphRAG summary and Order 2 (O2):
GraphRAG appears before RAG. We compare the
proportion of selected best samples from RAG and
GraphRAG, where a higher proportion indicates
better performance as predicted by the LLM.
The results of RAG vs. GraphRAG (Local) and
RAG vs. GraphRAG (Global) on the QMSum and
ODSum-story datasets are presented in Figure 4.
More result can be found in Appendix A.6. We
can make the following observations: (1) Posi**tion bias (Shi et al., 2024; Wang et al., 2024) is**
**evident in the LLM-as-a-Judge evaluations for**
**summarization task, as changing the order of the**
two methods significantly affects the predictions.
This effect is particularly strong in the comparison between RAG and GraphRAG (Local), where
the LLMs make completely opposite decisions
depending on the order, as shown in Figures 4a
and 4c. However, (2) Comparison between RAG
and GraphRAG (Global): While the proportions
vary, RAG consistently outperforms GraphRAG
(Global) in Comprehensiveness but underperforms
in Diversity as shown in Figures 4b and 4d. This result suggests that Community-based GraphRAG
**with Global Search focuses more on the global**
**aspects of whole corpus, whereas RAG captures**
**more detailed information.**
### 6 Conclusion
In this paper, we systematically evaluate and compare RAG and GraphRAG on general text-based
tasks. Our analysis reveals the distinct strengths
of RAG and GraphRAG in QA and query-based
summarization, as well as evaluation challenges in
summarization tasks, providing valuable insights
for future research. Building on these findings, we
propose two strategies to enhance QA performance.
Future work can explore improving GraphRAG
through better graph construction or developing
novel approaches to combine RAG and GraphRAG
methods for both effectiveness and efficiency.
-----
### Limitations
In this paper, we evaluate and compare RAG and
GraphRAG on Question Answering and Querybased Summarization tasks. Future work can extend this study to additional tasks to further assess
the strengths and applicability of GraphRAG. Additionally, the graph construction in all GraphRAG
methods explored in this work relies on LLM-based
construction, where LLMs extract entities and relations. However, other graph construction models
designed for text processing exist and can be investigated in future studies. Finally, we primarily
evaluate generation performance using Llama 3.18B-Instruct and Llama 3.1-70B-Instruct. Future
research can explore other generation models for a
more comprehensive comparison.
### References
Jiawei Chen, Hongyu Lin, Xianpei Han, and Le Sun.
2024. Benchmarking large language models in
retrieval-augmented generation. In Proceedings of
_the AAAI Conference on Artificial Intelligence, vol-_
ume 38, pages 17754–17762.
Jialin Dong, Bahare Fatemi, Bryan Perozzi, Lin F Yang,
and Anton Tsitsulin. 2024. Don’t forget to connect!
improving rag with graph-based reranking. arXiv
_preprint arXiv:2405.18414._
Qingxiu Dong, Lei Li, Damai Dai, Ce Zheng, Jingyuan
Ma, Rui Li, Heming Xia, Jingjing Xu, Zhiyong Wu,
Tianyu Liu, et al. 2022. A survey on in-context learning. arXiv preprint arXiv:2301.00234.
Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey,
Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman,
Akhil Mathur, Alan Schelten, Amy Yang, Angela
Fan, et al. 2024. The llama 3 herd of models. arXiv
_preprint arXiv:2407.21783._
Darren Edge, Ha Trinh, Newman Cheng, Joshua
Bradley, Alex Chao, Apurva Mody, Steven Truitt,
and Jonathan Larson. 2024. From local to global: A
graph rag approach to query-focused summarization.
_arXiv preprint arXiv:2404.16130._
Shahul Es, Jithin James, Luis Espinosa-Anke, and
Steven Schockaert. 2023. Ragas: Automated evaluation of retrieval augmented generation. _arXiv_
_preprint arXiv:2309.15217._
Wenqi Fan, Yujuan Ding, Liangbo Ning, Shijie Wang,
Hengyun Li, Dawei Yin, Tat-Seng Chua, and Qing
Li. 2024. A survey on rag meeting llms: Towards
retrieval-augmented large language models. In Pro_ceedings of the 30th ACM SIGKDD Conference on_
_Knowledge Discovery and Data Mining, pages 6491–_
6501.
Paulo Finardi, Leonardo Avila, Rodrigo Castaldoni, Pedro Gengo, Celio Larcher, Marcos Piau, Pablo Costa,
and Vinicius Caridá. 2024. The chronicles of rag:
The retriever, the chunk and the generator. arXiv
_preprint arXiv:2401.07883._
Yunfan Gao, Yun Xiong, Xinyu Gao, Kangxiang Jia,
Jinliu Pan, Yuxi Bi, Yi Dai, Jiawei Sun, and Haofen
Wang. 2023. Retrieval-augmented generation for
large language models: A survey. arXiv preprint
_arXiv:2312.10997._
Haoyu Han, Yu Wang, Harry Shomer, Kai Guo, Jiayuan
Ding, Yongjia Lei, Mahantesh Halappanavar, Ryan A
Rossi, Subhabrata Mukherjee, Xianfeng Tang, et al.
2024. Retrieval-augmented generation with graphs
(graphrag). arXiv preprint arXiv:2501.00309.
Haoyu Han, Yaochen Xie, Hui Liu, Xianfeng Tang,
Sreyashi Nag, William Headden, Yang Li, Chen Luo,
Shuiwang Ji, Qi He, et al. 2025. Reasoning with
graphs: Structuring implicit knowledge to enhance
llms reasoning. arXiv preprint arXiv:2501.07845.
Lei Huang, Weijiang Yu, Weitao Ma, Weihong Zhong,
Zhangyin Feng, Haotian Wang, Qianglong Chen,
Weihua Peng, Xiaocheng Feng, Bing Qin, et al. 2023.
A survey on hallucination in large language models:
Principles, taxonomy, challenges, and open questions.
_arXiv preprint arXiv:2311.05232._
Gautier Izacard, Patrick Lewis, Maria Lomeli, Lucas
Hosseini, Fabio Petroni, Timo Schick, Jane DwivediYu, Armand Joulin, Sebastian Riedel, and Edouard
Grave. 2023. Atlas: Few-shot learning with retrieval
augmented language models. Journal of Machine
_Learning Research, 24(251):1–43._
Zhengbao Jiang, Frank F Xu, Luyu Gao, Zhiqing
Sun, Qian Liu, Jane Dwivedi-Yu, Yiming Yang,
Jamie Callan, and Graham Neubig. 2023. Active retrieval augmented generation. arXiv preprint
_arXiv:2305.06983._
Vladimir Karpukhin, Barlas O˘guz, Sewon Min, Patrick
Lewis, Ledell Wu, Sergey Edunov, Danqi Chen, and
Wen-tau Yih. 2020. Dense passage retrieval for
open-domain question answering. arXiv preprint
_arXiv:2004.04906._
Jiho Kim, Sungjin Park, Yeonsu Kwon, Yohan Jo, James
Thorne, and Edward Choi. 2023. Factkg: Fact verification via reasoning on knowledge graphs. arXiv
_preprint arXiv:2305.06590._
Tom Kwiatkowski, Jennimaria Palomaki, Olivia Redfield, Michael Collins, Ankur Parikh, Chris Alberti,
Danielle Epstein, Illia Polosukhin, Jacob Devlin, Kenton Lee, et al. 2019. Natural questions: a benchmark
for question answering research. Transactions of the
_Association for Computational Linguistics, 7:453–_
466.
Xinze Li, Zhenghao Liu, Chenyan Xiong, Shi Yu,
Yu Gu, Zhiyuan Liu, and Ge Yu. 2023. Structureaware language model pretraining improves dense
-----
retrieval on structured data. _arXiv preprint_
_arXiv:2305.19912._
Yongqi Li, Wenjie Li, and Liqiang Nie. 2022. Dynamic
graph reasoning for conversational open-domain
question answering. ACM Transactions on Infor_mation Systems (TOIS), 40(4):1–24._
Chin-Yew Lin. 2004. Rouge: A package for automatic
evaluation of summaries. In Text summarization
_branches out, pages 74–81._
Fangru Lin, Emanuele La Malfa, Valentin Hofmann,
Elle Michelle Yang, Anthony Cohn, and Janet B
Pierrehumbert. 2024. Graph-enhanced large language models in asynchronous plan reasoning. arXiv
_preprint arXiv:2402.02805._
[Jerry Liu. 2022. LlamaIndex.](https://doi.org/10.5281/zenodo.1234)
Yi Liu, Lianzhe Huang, Shicheng Li, Sishuo Chen, Hao
Zhou, Fandong Meng, Jie Zhou, and Xu Sun. 2023.
Recall: A benchmark for llms robustness against
external counterfactual knowledge. arXiv preprint
_arXiv:2311.08147._
Xinbei Ma, Yeyun Gong, Pengcheng He, Hai Zhao,
and Nan Duan. 2023. Query rewriting for retrievalaugmented large language models. arXiv preprint
_arXiv:2305.14283._
Yao Ma and Jiliang Tang. 2021. _Deep learning on_
_graphs. Cambridge University Press._
Fatma Miladi, Valéry Psyché, and Daniel Lemire. 2024.
Leveraging gpt-4 for accuracy in education: A comparative study on retrieval-augmented generation in
moocs. In International Conference on Artificial
_Intelligence in Education, pages 427–434. Springer._
Zach Nussbaum, John X Morris, Brandon Duderstadt,
and Andriy Mulyar. 2024. Nomic embed: Training
a reproducible long context text embedder. arXiv
_preprint arXiv:2402.01613._
Boci Peng, Yun Zhu, Yongchao Liu, Xiaohe Bo,
Haizhou Shi, Chuntao Hong, Yan Zhang, and Siliang
Tang. 2024. Graph retrieval-augmented generation:
A survey. arXiv preprint arXiv:2408.08921.
Ori Ram, Yoav Levine, Itay Dalmedigos, Dor Muhlgay,
Amnon Shashua, Kevin Leyton-Brown, and Yoav
Shoham. 2023. In-context retrieval-augmented language models. Transactions of the Association for
_Computational Linguistics, 11:1316–1331._
Lin Shi, Chiyu Ma, Wenhua Liang, Weicheng Ma, and
Soroush Vosoughi. 2024. Judging the judges: A
systematic investigation of position bias in pairwise
comparative assessments by llms. arXiv preprint
_arXiv:2406.07791._
Yixuan Tang and Yi Yang. 2024. Multihop-rag: Benchmarking retrieval-augmented generation for multihop queries. arXiv preprint arXiv:2401.15391.
Yijun Tian, Huan Song, Zichen Wang, Haozhu Wang,
Ziqing Hu, Fang Wang, Nitesh V Chawla, and Panpan Xu. 2024. Graph neural prompting with large
language models. In Proceedings of the AAAI Con_ference on Artificial Intelligence, volume 38, pages_
19080–19088.
Alex Wang, Richard Yuanzhe Pang, Angelica Chen,
Jason Phang, and Samuel R Bowman. 2022. Squality:
Building a long-document summarization dataset the
hard way. arXiv preprint arXiv:2205.11465.
Zhen Wang. 2022. Modern question answering
datasets and benchmarks: A survey. arXiv preprint
_arXiv:2206.15030._
Ziqi Wang, Hanlin Zhang, Xiner Li, Kuan-Hao Huang,
Chi Han, Shuiwang Ji, Sham M Kakade, Hao Peng,
and Heng Ji. 2024. Eliminating position bias of
language models: A mechanistic approach. arXiv
_preprint arXiv:2407.01100._
Jerry Wei, Jason Wei, Yi Tay, Dustin Tran, Albert
Webson, Yifeng Lu, Xinyun Chen, Hanxiao Liu,
Da Huang, Denny Zhou, et al. 2023. Larger language
models do in-context learning differently. _arXiv_
_preprint arXiv:2303.03846._
Nirmalie Wiratunga, Ramitha Abeyratne, Lasal Jayawardena, Kyle Martin, Stewart Massie, Ikechukwu NkisiOrji, Ruvan Weerasinghe, Anne Liret, and Bruno
Fleisch. 2024. Cbr-rag: case-based reasoning for
retrieval augmented generation in llms for legal question answering. In International Conference on Case_Based Reasoning, pages 445–460. Springer._
Yaozu Wu, Yankai Chen, Zhishuai Yin, Weiping Ding,
and Irwin King. 2023. A survey on graph embedding techniques for biomedical data: Methods and
applications. Information Fusion, 100:101909.
Zonghan Wu, Shirui Pan, Fengwen Chen, Guodong
Long, Chengqi Zhang, and S Yu Philip. 2020. A comprehensive survey on graph neural networks. IEEE
_transactions on neural networks and learning sys-_
_tems, 32(1):4–24._
Feng Xia, Ke Sun, Shuo Yu, Abdul Aziz, Liangtian Wan,
Shirui Pan, and Huan Liu. 2021. Graph learning: A
survey. IEEE Transactions on Artificial Intelligence,
2(2):109–127.
Guangzhi Xiong, Qiao Jin, Zhiyong Lu, and
Aidong Zhang. 2024. Benchmarking retrievalaugmented generation for medicine. arXiv preprint
_arXiv:2402.13178._
Fangyuan Xu, Weijia Shi, and Eunsol Choi. 2023. Recomp: Improving retrieval-augmented lms with compression and selective augmentation. arXiv preprint
_arXiv:2310.04408._
Ran Xu, Wenqi Shi, Yue Yu, Yuchen Zhuang, Bowen
Jin, May D Wang, Joyce C Ho, and Carl Yang. 2024.
Ram-ehr: Retrieval augmentation meets clinical predictions on electronic health records. arXiv preprint
_arXiv:2403.00815._
-----
Shi-Qi Yan, Jia-Chen Gu, Yun Zhu, and Zhen-Hua Ling.
2024. Corrective retrieval augmented generation.
_arXiv preprint arXiv:2401.15884._
Zhilin Yang, Peng Qi, Saizheng Zhang, Yoshua Bengio, William W Cohen, Ruslan Salakhutdinov, and
Christopher D Manning. 2018. Hotpotqa: A dataset
for diverse, explainable multi-hop question answering. arXiv preprint arXiv:1809.09600.
Michihiro Yasunaga, Hongyu Ren, Antoine Bosselut, Percy Liang, and Jure Leskovec. 2021. Qagnn: Reasoning with language models and knowledge graphs for question answering. arXiv preprint
_arXiv:2104.06378._
Hao Yu, Aoran Gan, Kai Zhang, Shiwei Tong, Qi Liu,
and Zhaofeng Liu. 2024. Evaluation of retrievalaugmented generation: A survey. In CCF Conference
_on Big Data, pages 102–120. Springer._
Zichun Yu, Chenyan Xiong, Shi Yu, and Zhiyuan Liu.
2023. Augmentation-adapted retriever improves generalization of language models as generic plug-in.
_arXiv preprint arXiv:2305.17331._
Boyu Zhang, Hongyang Yang, Tianyu Zhou, Muhammad Ali Babar, and Xiao-Yang Liu. 2023. Enhancing
financial sentiment analysis via retrieval augmented
large language models. In Proceedings of the fourth
_ACM international conference on AI in finance, pages_
349–356.
Haozhen Zhang, Tao Feng, and Jiaxuan You. 2024.
Graph of records: Boosting retrieval augmented generation for long-context summarization with graphs.
_arXiv preprint arXiv:2410.11001._
Tianyi Zhang, Varsha Kishore, Felix Wu, Kilian Q
Weinberger, and Yoav Artzi. 2019. Bertscore: Evaluating text generation with bert. _arXiv preprint_
_arXiv:1904.09675._
Wayne Xin Zhao, Kun Zhou, Junyi Li, Tianyi Tang,
Xiaolei Wang, Yupeng Hou, Yingqian Min, Beichen
Zhang, Junjie Zhang, Zican Dong, et al. 2023. A
survey of large language models. _arXiv preprint_
_arXiv:2303.18223._
Huaixiu Steven Zheng, Swaroop Mishra, Xinyun Chen,
Heng-Tze Cheng, Ed H Chi, Quoc V Le, and Denny
Zhou. 2023a. Take a step back: Evoking reasoning via abstraction in large language models. arXiv
_preprint arXiv:2310.06117._
Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan
Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin,
Zhuohan Li, Dacheng Li, Eric Xing, et al. 2023b.
Judging llm-as-a-judge with mt-bench and chatbot
arena. Advances in Neural Information Processing
_Systems, 36:46595–46623._
Ming Zhong, Da Yin, Tao Yu, Ahmad Zaidi, Mutethia
Mutuma, Rahul Jha, Ahmed Hassan Awadallah, Asli
Celikyilmaz, Yang Liu, Xipeng Qiu, et al. 2021.
Qmsum: A new benchmark for query-based multidomain meeting summarization. _arXiv preprint_
_arXiv:2104.05938._
Yijie Zhou, Kejian Shi, Wencai Zhang, Yixin Liu, Yilun
Zhao, and Arman Cohan. 2023. Odsum: New benchmarks for open domain multi-document summarization. arXiv preprint arXiv:2309.08960.
-----
### A Appendix
**A.1** **Dataset**
In this section, we introduce the used datasets in the question answering tasks and query-based summarization tasks.
**A.1.1** **Question Answering**
In the QA tasks, we use the following four widely used datasets:
- Natural Questions (NQ) (Kwiatkowski et al., 2019): The NQ dataset is a widely used benchmark
for evaluating open-domain question answering systems. Introduced by Google, it consists of real
user queries from Google Search with corresponding answers extracted from Wikipedia. Since it
primarily contains single-hop questions, we use NQ as the representative dataset for single-hop
QA. We treat NQ as a single-document QA task, where multiple questions are associated with each
document. Accordingly, we build a separate RAG system for each document in the dataset.
- Hotpot (Yang et al., 2018): HotpotQA is a widely used multi-hop question dataset that provides
10 paragraphs per question. The dataset includes varying difficulty levels, with easier questions
often solvable by LLMs. To ensure a more challenging evaluation, we randomly selected 1,000 hard
bridging questions from the development set of HotpotQA. Additionally, we treat HotpotQA as a
multi-document QA task and build a single RAG system to handle all questions.
- MultiHop-RAG (Tang and Yang, 2024): MultiHop-RAG is a QA dataset designed to evaluate
retrieval and reasoning across multiple documents with metadata in RAG pipelines. Constructed
from English news articles, it contains 2,556 queries, with supporting evidence distributed across 2
to 4 documents. The dataset includes four query types: Inference queries, which synthesize claims
about a bridge entity to identify it; Comparison queries, which compare similarities or differences
and typically yield "yes" or "no" answers; Temporal queries, which examine event ordering with
answers like "before" or "after"; and Null queries, where no answer can be derived from the retrieved
documents. It is also a multi-document QA task.
- NovelQA (Tang and Yang, 2024): NovelQA is a benchmark designed to evaluate the long-text
understanding and retrieval ability of LLMs using manually curated questions about English novels
exceeding 50,000 words. The dataset includes queries that focus on minor details or require crosschapter reasoning, making them inherently challenging for LLMs. It covers various query types
such as details, multi-hop, single-hop, character, meaning, plot, relation, setting, span, and times.
Key challenges highlighted by NovelQA include grasping abstract meanings (meaning questions),
understanding nuanced relationships (relation questions), and tracking temporal sequences and spatial
extents (span and time questions), emphasizing the difficulty of maintaining and applying contextual
information across long narratives. We use it for single-document QA task.
**A.1.2** **Query-based Summarization**
In the Query-based Summarization tasks, we adopt the following four widely used datasets:
- SQuALITY (Wang et al., 2022): SQuALITY (Summary-format QUestion Answering with Long
Input Texts) is a question-focused, long-document, multi-reference summarization dataset. It consists
of short stories from Project Gutenberg, each ranging from 4,000 to 6,000 words. Each story is paired
with five questions, and each question has four reference summaries written by Upwork writers and
NYU undergraduates. SQuALITY is designed as a single-document summarization task, making it a
valuable benchmark for evaluating summarization models on long-form content.
- QMSum (Zhong et al., 2021): QMSum is a human-annotated benchmark for query-based, multidomain meeting summarization, containing 1,808 query-summary pairs from 232 meetings across
multiple domains. We use QMSum as a single-document summarization task in our evaluation.
-----
- ODSum (Zhou et al., 2023): The ODSum dataset is designed to evaluate modern summarization
models in multi-document contexts and consists of two subsets: ODSum-story and ODSum-meeting.
ODSum-story is derived from the SQuALITY dataset, while ODSum-meeting is constructed from
QMSum. We use both ODSum-story and ODSum-meeting for the multi-document summarization
task in our evaluation.
**A.2** **More results on NovelQA dataset**
In this section, we present the missing results for the NovelQA dataset from the main sections. These include the performance of KG-GraphRAG (Triplets) with LLaMA 3.1-8B (Table 6), RAG with LLaMA 3.170B (Table 7), KG-GraphRAG (Triplets) with LLaMA 3.1-70B (Table 8), KG-GraphRAG (Triplets+Text)
with LLaMA 3.1-70B (Table 9), Community-GraphRAG (Local) with LLaMA 3.1-70B (Table 10), and
Community-GraphRAG (Global) with LLaMA 3.1-70B (Table 11).
Table 6: The performance of KG-GraphRAG (Triplets) with Llama 3.1-8B model on NovelQA dataset.
KG-GraphRAG(Triplet) character meaning plot relat settg span times avg
mh 31.25 17.65 41.67 50.56 38.46 64 26.47 32.89
sh 35.53 45.71 30.54 62.5 27.84 - - 33.75
dtl 31.43 24.72 35.71 17.86 27.03 - - 27.37
avg 33.7 29.81 32.63 44 28.57 64 26.47 31.88
Table 7: The performance of RAG with Llama 3.1-70B model on NovelQA dataset.
RAG character meaning plot relat settg span times avg
mh 64.58 82.35 77.78 69.66 84.62 36 36.63 48.5
sh 70.39 70 76.57 75 83.51 - - 75.27
dtl 60 51.12 76.79 67.86 83.78 - - 61.25
avg 66.67 58.11 76.74 69.6 83.67 36 36.63 61.42
Table 8: The performance of KG-GraphRAG (Triplets) with Llama 3.1-70B model on NovelQA dataset.
KG-GraphRAG (Triplets) character meaning plot relat settg span times avg
mh 50 76.47 75 43.82 76.92 24 22.46 33.72
sh 52.63 62.86 55.23 12.5 50.52 - - 54.06
dtl 35.71 26.97 39.29 53.57 37.84 - - 33.6
avg 47.78 39.62 54.68 44 49.66 24 22.46 41.18
Table 9: The performance of KG-GraphRAG (Triplets+Text) with Llama 3.1-70B model on NovelQA dataset.
KG-GraphRAG (Triplets+Text) character meaning plot relat settg span times avg
mh 56.25 58.82 63.89 51.69 84.62 24 21.39 33.72
sh 51.97 61.43 55.65 50 50.52 - - 54.42
dtl 34.29 25.28 41.07 50 37.84 - - 32.52
avg 48.15 36.98 54.08 51.2 50.34 24 21.39 41.05
**A.3** **RAG vs. GraphRAG Selection**
We classify QA queries into Fact-based and Reasoning-based queries. Fact-based queries are processed
using RAG, while Reasoning-based queries are handled by GraphRAG. The Query Classification prompt
is shown in Figure 5.
|KG-GraphRAG(Triplet)|character meaning plot relat settg span times avg|
|---|---|
|mh sh dtl avg|31.25 17.65 41.67 50.56 38.46 64 26.47 32.89 35.53 45.71 30.54 62.5 27.84 - - 33.75 31.43 24.72 35.71 17.86 27.03 - - 27.37 33.7 29.81 32.63 44 28.57 64 26.47 31.88|
|RAG|character meaning plot relat settg span times avg|
|---|---|
|mh sh dtl avg|64.58 82.35 77.78 69.66 84.62 36 36.63 48.5 70.39 70 76.57 75 83.51 - - 75.27 60 51.12 76.79 67.86 83.78 - - 61.25 66.67 58.11 76.74 69.6 83.67 36 36.63 61.42|
|KG-GraphRAG (Triplets)|character meaning plot relat settg span times avg|
|---|---|
|mh sh dtl avg|50 76.47 75 43.82 76.92 24 22.46 33.72 52.63 62.86 55.23 12.5 50.52 - - 54.06 35.71 26.97 39.29 53.57 37.84 - - 33.6 47.78 39.62 54.68 44 49.66 24 22.46 41.18|
|KG-GraphRAG (Triplets+Text)|character meaning plot relat settg span times avg|
|---|---|
|mh sh dtl avg|56.25 58.82 63.89 51.69 84.62 24 21.39 33.72 51.97 61.43 55.65 50 50.52 - - 54.42 34.29 25.28 41.07 50 37.84 - - 32.52 48.15 36.98 54.08 51.2 50.34 24 21.39 41.05|
-----
Table 10: The performance of Community-GraphRAG (Local) with Llama 3.1-70B model on NovelQA dataset.
Community-GraphRAG (Local) character meaning plot relat settg span times avg
mh 77.08 70.59 63.89 77.53 92.31 28 32.35 46.68
sh 68.42 71.43 74.9 62.5 74.23 - - 72.44
dtl 55.71 37.08 69.64 64.29 75.68 - - 51.49
avg 66.67 48.3 72.81 73.6 76.19 28 32.35 57.32
Table 11: The performance of Community-GraphRAG (Global) with Llama 3.1-70B model on NovelQA dataset.
Community-GraphRAG (Global) character meaning plot relat settg span times avg
mh 47.92 58.82 55.56 57.3 61.54 16 35.83 41.53
sh 42.76 42.86 54.39 25 40.21 - - 47
dtl 24.29 22.47 32.14 50 35.14 - - 27.64
avg 38.89 30.19 50.76 53.6 40.82 16 35.83 40.21
**Prompt for Query Classification**
System Prompt: Classifying Queries into Fact-Based and Reasoning-Based Categories
You are an AI model tasked with classifying queries into one of two categories based on their
complexity and reasoning requirements.
**Category Definitions**
1. Fact-Based Queries
- The answer can be directly retrieved from a knowledge source or requires details.
- The query does not require multi-step reasoning, inference, or cross-referencing multiple sources.
2. Reasoning-Based Queries
- The answer cannot be found in a single lookup and requires cross-referencing multiple sources,
logical inference, or multi-step reasoning.
**Examples**
**Fact-Based Queries**
{{ Fact-Based Queries Examples }}
**Reasoning-Based Queries**
{{ Reasoning-Based Queries Examples }}
Figure 5: Prompt for Query Classification.
**A.4** **Query-based Summarization Results with Llama3.1-70B model**
In this section, we present the results for Query-based Summarization tasks using the LLaMA 3.1-70B
model. The results for single-document summarization are shown in Table 12, while the results for
multi-document summarization are provided in Table 13.
Table 12: The performance of query-based single document summarization task using Llama3.1-70B.
|Community-GraphRAG (Local)|character meaning plot relat settg span times avg|
|---|---|
|mh sh dtl avg|77.08 70.59 63.89 77.53 92.31 28 32.35 46.68 68.42 71.43 74.9 62.5 74.23 - - 72.44 55.71 37.08 69.64 64.29 75.68 - - 51.49 66.67 48.3 72.81 73.6 76.19 28 32.35 57.32|
|Community-GraphRAG (Global)|character meaning plot relat settg span times avg|
|---|---|
|mh sh dtl avg|47.92 58.82 55.56 57.3 61.54 16 35.83 41.53 42.76 42.86 54.39 25 40.21 - - 47 24.29 22.47 32.14 50 35.14 - - 27.64 38.89 30.19 50.76 53.6 40.82 16 35.83 40.21|
|SQuALITY QMSum Method ROUGE-2 BERTScore ROUGE-2 BERTScore P R F1 P R F1 P R F1 P R F1|SQuALITY|QMSum|
|---|---|---|
||ROUGE-2 BERTScore|ROUGE-2 BERTScore|
||P R F1 P R F1|P R F1 P R F1|
|RAG KG-GraphRAG(Triplets only) KG-GraphRAG(Triplets+Text) Community-GraphRAG(Local) Community-GraphRAG(Global) Combine|11.85 14.24 11.00 85.96 85.76 85.67 8.53 10.28 7.46 84.13 83.97 83.89 6.57 10.14 6.00 80.52 82.23 81.07 12.54 10.31 9.61 84.50 85.33 84.71 8.99 4.78 5.60 81.64 83.64 82.44 13.59 11.32 10.55 84.88 85.76 85.12|10.42 10.00 9.53 86.14 85.92 86.02 10.62 6.25 7.48 83.20 84.72 83.94 8.64 7.85 7.29 84.10 84.55 84.31 13.69 7.43 9.14 84.09 85.85 84.95 10.97 4.40 6.01 81.93 84.67 83.26 13.16 8.67 9.93 85.18 86.21 85.69|
-----
Table 13: The performance of query-based multiple document summarization task using Llama3.1-70B.
ODSum-story ODSum-meeting
Method ROUGE-2 BERTScore ROUGE-2 BERTScore
P R F1 P R F1 P R F1 P R
RAG 15.60 9.98 11.09 74.80 81.29 77.89 18.81 6.41 8.97 83.56 85.16
KG-GraphRAG(Triplets only) 10.08 9.12 8.48 75.71 81.93 78.66 11.52 3.41 4.79 81.19 83.07
KG-GraphRAG(Triplets+Text) 10.98 16.67 11.42 76.74 81.92 79.21 13.09 6.31 7.70 84.07 84.24
Community-GraphRAG(Local) 14.20 11.34 11.25 75.44 81.81 78.46 16.17 7.87 9.23 84.17 84.85
Community-GraphRAG(Global) 10.46 6.30 7.08 74.63 81.24 77.77 10.65 1.99 3.28 79.78 82.53
Combine 14.76 12.17 11.72 75.39 81.75 78.41 17.57 8.64 10.34 84.51 85.14
**A.5** **The LLM-as-a-Judge Prompt**
The LLM-as-a-Judge prompt can be found in Figure 6.
|ODSum-story ODSum-meeting Method ROUGE-2 BERTScore ROUGE-2 BERTScore P R F1 P R F1 P R F1 P R F1|ODSum-story|ODSum-meeting|
|---|---|---|
||ROUGE-2 BERTScore|ROUGE-2 BERTScore|
||P R F1 P R F1|P R F1 P R F1|
|RAG KG-GraphRAG(Triplets only) KG-GraphRAG(Triplets+Text) Community-GraphRAG(Local) Community-GraphRAG(Global) Combine|15.60 9.98 11.09 74.80 81.29 77.89 10.08 9.12 8.48 75.71 81.93 78.66 10.98 16.67 11.42 76.74 81.92 79.21 14.20 11.34 11.25 75.44 81.81 78.46 10.46 6.30 7.08 74.63 81.24 77.77 14.76 12.17 11.72 75.39 81.75 78.41|18.81 6.41 8.97 83.56 85.16 84.34 11.52 3.41 4.79 81.19 83.07 82.11 13.09 6.31 7.70 84.07 84.24 84.14 16.17 7.87 9.23 84.17 84.85 84.49 10.65 1.99 3.28 79.78 82.53 81.12 17.57 8.64 10.34 84.51 85.14 84.81|
**LLM-as-a-Judge Prompt**
You are an expert evaluator assessing the quality of responses in a query-based summarization task.
Below is a query, followed by two LLM-generated summarization answers. Your task is to evaluate
the best answer based on the given criteria. For each aspect, select the model that performs better.
**Query**
{{query}}
**Answers Section**
**The Answer of Model 1:**
{{answer 1}}
**The Answer of Model 2:**
{{answer 2}}
**Evaluation Criteria Assess each LLM-generated answer independently based on the following**
two aspects:
1. Comprehensiveness
- Does the answer fully address the query and include all relevant information?
- A comprehensive answer should cover all key points, ensuring that no important details are
missing.
- It should present a well-rounded view, incorporating relevant context when necessary.
- The level of detail should be sufficient to fully inform the reader without unnecessary omission
or excessive brevity.
2. Global Diversity
- Does the answer provide a broad and globally inclusive perspective?
- A globally diverse response should avoid narrow or region-specific biases and instead consider
multiple viewpoints.
- The response should be accessible and relevant to a wide, international audience rather than
assuming familiarity with specific local contexts.
Figure 6: LLM-as-a-Judge Prompt.
**A.6** **The LLM-as-a-Judge Results on more datasets**
In the main section, we present LLM-as-a-Judge results for the OMSum and ODSum-story datasets. Here,
we provide additional results on the SQuALITY and ODSum-meeting datasets, as shown in Figure 7.
-----
1.0
0.8
1.0
0.8
1.0
0.8
1.0
0.8
0.6
0.4
0.6
0.4
0.6
0.4
0.6
0.4
0.2
0.0
Comprehensiveness Diversity
0.2
0.0
Comprehensiveness Diversity
0.2
0.0
Comprehensiveness Diversity
0.2
0.0
Comprehensiveness Diversity
|RAG-Order 1 GraphRAG-Local-Order 1 RAG-Order 2 GraphRAG-Local-Order 2|RAG-Order 1 GraphRAG-Local-Order 1 RAG-Order 2 GraphRAG-Local-Order 2|Col3|Col4|Col5|Col6|
|---|---|---|---|---|---|
||||-Loc|||
|||||||
|||||||
|||||||
|||||||
|Col1|Col2|Col3|Col4|
|---|---|---|---|
|||||
|||||
|||||
|RAG-Order 1 GraphRAG-Gloabl-Order 1 RAG-Order 2 GraphRAG-Gloabl-Order 2|RAG-Order 1 GraphRAG-Gloabl-Order 1 RAG-Order 2 GraphRAG-Gloabl-Order 2|Col3|Col4|
|---|---|---|---|
|||||
|||||
|Col1|Col2|Col3|Col4|
|---|---|---|---|
|||||
|||||
|||||
|RAG-Order 1 GraphRAG-Local-Order 1 RAG-Order 2 GraphRAG-Local-Order 2|Col2|Col3|Col4|RAG-Order 1 GraphRAG-Local-Order 1 RAG-Order 2 GraphRAG-Local-Order 2|Col6|Col7|Col8|Col9|Col10|Col11|
|---|---|---|---|---|---|---|---|---|---|---|
||||||R G||||||
||||||||||||
||||||||||||
||||||||||||
||||||||||||
||||||||||||
|RAG-Order 1 GraphRAG-Gloabl-Order 1 RAG-Order 2 GraphRAG-Gloabl-Order 2|Col2|Col3|RAG-Order 1 GraphRAG-Gloabl-Order 1 RAG-Order 2 GraphRAG-Gloabl-Order 2|Col5|Col6|Col7|Col8|Col9|Col10|
|---|---|---|---|---|---|---|---|---|---|
|||||||||||
||||||||Gloa||der 2|
|||||||||||
|||||||||||
|||||||||||
|||||||||||
(a) SQuALITY Local
(b) SQuALITY Global
(c) ODSum-meeting Local
(d) ODSum-meeting Global
Figure 7: Comparison of LLM-as-a-Judge evaluations for RAG and GraphRAG. "Local" refers to the evaluation of
RAG vs. GraphRAG-Local, while "Global" refers to RAG vs. GraphRAG-Global. "Order 1" corresponds to the
prompt where RAG result is presented before GraphRAG, whereas "Order 2" corresponds to the reversed order.
-----
|
{
"id": "2502.11371",
"submitter": "Haoyu Han",
"authors": "Haoyu Han, Harry Shomer, Yu Wang, Yongjia Lei, Kai Guo, Zhigang Hua,\n Bo Long, Hui Liu, Jiliang Tang",
"title": "RAG vs. GraphRAG: A Systematic Evaluation and Key Insights",
"comments": null,
"journal-ref": null,
"doi": null,
"report-no": null,
"categories": "cs.IR",
"license": "http://creativecommons.org/licenses/by/4.0/",
"abstract": " Retrieval-Augmented Generation (RAG) enhances the performance of LLMs across\nvarious tasks by retrieving relevant information from external sources,\nparticularly on text-based data. For structured data, such as knowledge graphs,\nGraphRAG has been widely used to retrieve relevant information. However, recent\nstudies have revealed that structuring implicit knowledge from text into graphs\ncan benefit certain tasks, extending the application of GraphRAG from graph\ndata to general text-based data. Despite their successful extensions, most\napplications of GraphRAG for text data have been designed for specific tasks\nand datasets, lacking a systematic evaluation and comparison between RAG and\nGraphRAG on widely used text-based benchmarks. In this paper, we systematically\nevaluate RAG and GraphRAG on well-established benchmark tasks, such as Question\nAnswering and Query-based Summarization. Our results highlight the distinct\nstrengths of RAG and GraphRAG across different tasks and evaluation\nperspectives. Inspired by these observations, we investigate strategies to\nintegrate their strengths to improve downstream tasks. Additionally, we provide\nan in-depth discussion of the shortcomings of current GraphRAG approaches and\noutline directions for future research.\n",
"versions": {
"version": [
"v1"
],
"created": [
"Mon, 17 Feb 2025 02:36:30 GMT"
]
},
"update_date": "2025-02-18",
"authors_parsed": [
[
"Han",
"Haoyu",
""
],
[
"Shomer",
"Harry",
""
],
[
"Wang",
"Yu",
""
],
[
"Lei",
"Yongjia",
""
],
[
"Guo",
"Kai",
""
],
[
"Hua",
"Zhigang",
""
],
[
"Long",
"Bo",
""
],
[
"Liu",
"Hui",
""
],
[
"Tang",
"Jiliang",
""
]
]
}
|
## LIGHTRAG: SIMPLE AND FAST RETRIEVAL-AUGMENTED GENERATION
**Zirui Guo[1][,][2], Lianghao Xia[2], Yanhua Yu[1], Tu Ao[1], Chao Huang[2][∗]**
Beijing University of Posts and Telecommunications[1]
University of Hong Kong[2]
zrguo101@hku.hk aka_xia@foxmail.com chaohuang75@gmail.com
#### ABSTRACT
Retrieval-Augmented Generation (RAG) systems enhance large language models
(LLMs) by integrating external knowledge sources, enabling more accurate and
contextually relevant responses tailored to user needs. However, existing RAG
systems have significant limitations, including reliance on flat data representations and inadequate contextual awareness, which can lead to fragmented answers
that fail to capture complex inter-dependencies. To address these challenges, we
propose LightRAG, which incorporates graph structures into text indexing and
retrieval processes. This innovative framework employs a dual-level retrieval system that enhances comprehensive information retrieval from both low-level and
high-level knowledge discovery. Additionally, the integration of graph structures
with vector representations facilitates efficient retrieval of related entities and their
relationships, significantly improving response times while maintaining contextual
relevance. This capability is further enhanced by an incremental update algorithm
that ensures the timely integration of new data, allowing the system to remain
effective and responsive in rapidly changing data environments. Extensive experimental validation demonstrates considerable improvements in retrieval accuracy
and efficiency compared to existing approaches. We have made our LightRAG
[open-source and available at the link: https://github.com/HKUDS/LightRAG.](https://github.com/HKUDS/LightRAG)
#### 1 INTRODUCTION
Retrieval-Augmented Generation (RAG) systems have been developed to enhance large language
models (LLMs) by integrating external knowledge sources Sudhi et al. (2024); Es et al. (2024);
Salemi & Zamani (2024). This innovative integration allows LLMs to generate more accurate and
contextually relevant responses, significantly improving their utility in real-world applications. By
adapting to specific domain knowledge Tu et al. (2024), RAG systems ensure that the information
provided is not only pertinent but also tailored to the user’s needs. Furthermore, they offer access to
up-to-date information Zhao et al. (2024), which is crucial in rapidly evolving fields. Chunking plays
a vital role in facilitating the retrieval-augmented generation process Lyu et al. (2024). By breaking
down a large external text corpus into smaller, more manageable segments, chunking significantly
enhances the accuracy of information retrieval. This approach allows for more targeted similarity
searches, ensuring that the retrieved content is directly relevant to user queries.
However, existing RAG systems have key limitations that hinder their performance. First, many
methods rely on flat data representations, restricting their ability to understand and retrieve information
based on intricate relationships between entities. Second, these systems often lack the contextual
awareness needed to maintain coherence across various entities and their interrelations, resulting
in responses that may not fully address user queries. For example, consider a user asking, “How
does the rise of electric vehicles influence urban air quality and public transportation infrastructure?”
Existing RAG methods might retrieve separate documents on electric vehicles, air pollution, and
public transportation challenges but struggle to synthesize this information into a cohesive response.
They may fail to explain how the adoption of electric vehicles can improve air quality, which in turn
_∗Chao Huang is the corresponding author._
-----
could affect public transportation planning. As a result, the user may receive a fragmented answer
that does not adequately capture the complex inter-dependencies among these topics.
To address these limitations, we propose incorporating graph structures into text indexing and
relevant information retrieval. Graphs are particularly effective at representing the interdependencies
among different entities Rampášek et al. (2022), which enables a more nuanced understanding
of relationships. The integration of graph-based knowledge structures facilitates the synthesis of
information from multiple sources into coherent and contextually rich responses. Despite these
advantages, developing a fast and scalable graph-empowered RAG system that efficiently handles
varying query volumes is crucial. In this work, we achieve an effective and efficient RAG system by
addressing three key challenges: i) Comprehensive Information Retrieval. Ensuring comprehensive
information retrieval that captures the full context of inter-dependent entities from all documents;
ii) Enhanced Retrieval Efficiency. Improving retrieval efficiency over the graph-based knowledge
structures to significantly reduce response times; iii) Rapid Adaptation to New Data. Enabling
quick adaptation to new data updates, ensuring the system remains relevant in dynamic environments.
In response to the outlined challenges, we propose LightRAG, a model that seamlessly integrates a
graph-based text indexing paradigm with a dual-level retrieval framework. This innovative approach
enhances the system’s capacity to capture complex inter-dependencies among entities, resulting in
more coherent and contextually rich responses. LightRAG employs efficient dual-level retrieval
strategies: low-level retrieval, which focuses on precise information about specific entities and their
relationships, and high-level retrieval, which encompasses broader topics and themes. By combining
both detailed and conceptual retrieval, LightRAG effectively accommodates a diverse range of quries,
ensuring that users receive relevant and comprehensive responses tailored to their specific needs.
Additionally, by integrating graph structures with vector representations, our framework facilitates
efficient retrieval of related entities and relations while enhancing the comprehensiveness of results
through relevant structural information from the constructed knowledge graph.
In summary, the key contributions of this work are highlighted as follows:
- General Aspect. We emphasize the importance of developing a graph-empowered RAG system to
overcome the limitations of existing methods. By integrating graph structures into text indexing,
we can effectively represent complex interdependencies among entities, fostering a nuanced
understanding of relationships and enabling coherent, contextually rich responses.
- Methodologies. To enable an efficient and adaptive RAG system, we propose LightRAG, which
integrates a dual-level retrieval paradigm with graph-enhanced text indexing. This approach
captures both low-level and high-level information for comprehensive, cost-effective retrieval.
By eliminating the need to rebuild the entire index, LightRAG reduces computational costs and
accelerates adaptation, while its incremental update algorithm ensures timely integration of new
data, maintaining effectiveness in dynamic environments.
- Experimental Findings. Extensive experiments were conducted to evaluate the effectiveness of
LightRAG in comparison to existing RAG models. These assessments focused on several key
dimensions, including retrieval accuracy, model ablation, response efficiency, and adaptability to
new information. The results demonstrated significant improvements over baseline methods.
#### 2 RETRIEVAL-AUGMENTED GENERATION
Retrieval-Augmented Generation (RAG) integrates user queries with a collection of pertinent documents sourced from an external knowledge database, incorporating two essential elements: the
**Retrieval Component and the Generation Component. 1) The retrieval component is responsible**
for fetching relevant documents or information from the external knowledge database. It identifies and
retrieves the most pertinent data based on the input query. 2) After the retrieval process, the generation
component takes the retrieved information and generates coherent, contextually relevant responses. It
leverages the capabilities of the language model to produce meaningful outputs. Formally, this RAG
framework, denoted as M, can be defined as follows:
_M =_ �G, R = (φ, ψ)�, _M(q; D) = G�q, ψ(q; D[ˆ])�,_ _Dˆ = φ(D)_ (1)
In this framework, G and R represent the generation module and the retrieval module, respectively,
while q denotes the input query and D refers to the external database. The retrieval module R
-----
...the methods and strategies employedbycolonies and ensure their health andproductivity. A Beekeeper is anindividual who produces honey andother related products, playing acrucial role in ……Beekeepers includingmonitoringpreventingutilizing techniques to handle bees,such as using smoke to calm them ... BEEKEEPERbeekeepers observing bee engage in various tasks,pesthive‘s practices involvetoinfestations,manageconditions,behavior,andbee Entity & Rel ExtractionBeekeeperBeekeeperBeekeeperLLM ProfilingDeduplicationMatchObserve…A beekeeper is an person beekeeperwho…Bees used for RetrievalIndex Graph **SourceTargetKeywordsDescription:negatively impacted ...Original Chunks IDEntity NameEntity TypeDescriptionan individual who produces ...Original Chunks ID: Industrial agriculture: Honey Bee: Agriculture ...: A Beekeeper is Honey Bees are : PERSON: Beekeeper: xxx: xxx** BeekeeperEnvironmental ImpactAgricultureFormersLow-level KeysHigh-level KeysQuery + LLMHoney BeeProductionHive **… …** Retrieved ContentOriginal TextRelationsEntities "Beekeeper",”individual who produces honey andother related products, playing acrucial role in ......"Beekeeper",“bee","Beekeepersmanage bees but do not developindividual relationships with themdue to the limited interaction timewith each hive.”BEEKEEPER's practices involve themethods and strategies employedby beekeepers to manage ……-----Entities------Relationships-----Contexts----A Beekeeper is an”
Graph-based Text Indexing Dual-level Retrieval Paradigm
|observing|bee|
|---|---|
Figure 1: Overall architecture of the proposed LightRAG framework.
includes two key functionalities: i) Data Indexer φ(·): which involves building a specific data
structure _D[ˆ] based on the external database D. ii) Data Retriever ψ(·): The relevant documents are_
obtained by comparing the query against the indexed data, also denoted as “relevant documents”. By
leveraging the information retrieved through ψ(·) along with the initial query q, the generative model
_G(·) efficiently produces high-quality, contextually relevant responses._
In this work, we target several key points essential for an efficient and effective Retrieval-Augmented
Generation (RAG) system which are elaborated below:
- Comprehensive Information Retrieval: The indexing function φ(·) must be adept at extracting
global information, as this is crucial for enhancing the model’s ability to answer queries effectively.
- Efficient and Low-Cost Retrieval: The indexed data structure _D[ˆ] must enable rapid and cost-_
efficient retrieval to effectively handle a high volume of queries.
- Fast Adaptation to Data Changes: The ability to swiftly and efficiently adjust the data structure
to incorporate new information from the external knowledge base, is crucial for ensuring that the
system remains current and relevant in an ever-changing information landscape.
#### 3 THE LIGHTRAG ARCHITECTURE
3.1 GRAPH-BASED TEXT INDEXING
**Graph-Enhanced Entity and Relationship Extraction. Our LightRAG enhances the retrieval**
system by segmenting documents into smaller, more manageable pieces. This strategy allows for
quick identification and access to relevant information without analyzing entire documents. Next,
we leverage LLMs to identify and extract various entities (e.g., names, dates, locations, and events)
along with the relationships between them. The information collected through this process will be
used to create a comprehensive knowledge graph that highlights the connections and insights across
the entire collection of documents. We formally represent this graph generation module as follows:
_Dˆ = (ˆV, ˆE) = Dedupe ◦_ Prof(V, E), _V, E = ∪Di∈DRecog(Di)_ (2)
where _D[ˆ] represents the resulting knowledge graphs. To generate this data, we apply three main_
processing steps to the raw text documents Di. These steps utilize a LLM for text analysis and
processing. Details about the prompt templates and specific settings for this part can be found in
Appendix 7.3.2. The functions used in our graph-based text indexing paradigm are described as:
- Extracting Entities and Relationships. R(·): This function prompts a LLM to identify entities
(nodes) and their relationships (edges) within the text data. For instance, it can extract entities
like "Cardiologists" and "Heart Disease," and relationships such as "Cardiologists diagnose Heart
Disease" from the text: "Cardiologists assess symptoms to identify potential heart issues." To
improve efficiency, the raw text D is segmented into multiple chunks Di.
- LLM Profiling for Key-Value Pair Generation. P(·): We employ a LLM-empowered profiling
function, P(·), to generate a text key-value pair (K, V ) for each entity node in V and relation
edge in E. Each index key is a word or short phrase that enables efficient retrieval, while the
corresponding value is a text paragraph summarizing relevant snippets from external data to aid in
text generation. Entities use their names as the sole index key, whereas relations may have multiple
index keys derived from LLM enhancements that include global themes from connected entities.
- Deduplication to Optimize Graph Operations. D(·): Finally, we implement a deduplication
function, D(·), that identifies and merges identical entities and relations from different segments of
-----
the raw text Di. This process effectively reduces the overhead associated with graph operations on
_Dˆ by minimizing the graph’s size, leading to more efficient data processing._
Our LightRAG offers two advantages through its graph-based text indexing paradigm. First, Com**prehensive Information Understanding. The constructed graph structures enable the extraction**
of global information from multi-hop subgraphs, greatly enhancing LightRAG’s ability to handle
complex queries that span multiple document chunks. Second, Enhanced Retrieval Performance.
the key-value data structures derived from the graph are optimized for rapid and precise retrieval.
This provides a superior alternative to less accurate embedding matching methods (Gao et al., 2023)
and inefficient chunk traversal techniques (Edge et al., 2024) commonly used in existing approaches.
**Fast Adaptation to Incremental Knowledge Base. To efficiently adapt to evolving data changes**
while ensuring accurate and relevant responses, our LightRAG incrementally updates the knowledge
base without the need for complete reprocessing of the entire external database. For a new document
_D[′], the incremental update algorithm processes it using the same graph-based indexing steps φ as_
before, resulting in _D[ˆ][′]_ = (V[ˆ][′], _E[ˆ][′]). Subsequently, LightRAGcombines the new graph data with the_
original by taking the union of the node sets _V[ˆ] and_ _V[ˆ][′], as well as the edge sets_ _E[ˆ] and_ _E[ˆ][′]._
Two key objectives guide our approach to fast adaptation for the incremental knowledge base: Seam**less Integration of New Data. By applying a consistent methodology to new information, the**
incremental update module allows the LightRAG to integrate new external databases without disrupting the existing graph structure. This approach preserves the integrity of established connections,
ensuring that historical data remains accessible while enriching the graph without conflicts or redundancies. Reducing Computational Overhead . By eliminating the need to rebuild the entire index
graph, this method reduces computational overhead and facilitates the rapid assimilation of new data.
Consequently, LightRAG maintains system accuracy, provides current information, and conserves
resources, ensuring users receive timely updates and enhancing the overall RAG effectiveness.
3.2 DUAL-LEVEL RETRIEVAL PARADIGM
To retrieve relevant information from both specific document chunks and their complex interdependencies, our LightRAG proposes generating query keys at both detailed and abstract levels.
- Specific Queries. These queries are detail-oriented and typically reference specific entities within
the graph, requiring precise retrieval of information associated with particular nodes or edges. For
example, a specific query might be, “Who wrote ’Pride and Prejudice’?”
- Abstract Queries. In contrast, abstract queries are more conceptual, encompassing broader topics,
summaries, or overarching themes that are not directly tied to specific entities. An example of an
abstract query is, “How does artificial intelligence influence modern education?”
To accommodate diverse query types, the LightRAG employs two distinct retrieval strategies within
the dual-level retrieval paradigm. This ensures that both specific and abstract inquiries are addressed
effectively, allowing the system to deliver relevant responses tailored to user needs.
- Low-Level Retrieval. This level is primarily focused on retrieving specific entities along with their
associated attributes or relationships. Queries at this level are detail-oriented and aim to extract
precise information about particular nodes or edges within the graph.
- High-Level Retrieval. This level addresses broader topics and overarching themes. Queries at this
level aggregate information across multiple related entities and relationships, providing insights
into higher-level concepts and summaries rather than specific details.
**Integrating Graph and Vectors for Efficient Retrieval. By combining graph structures with**
vector representations, the model gains a deeper insight into the interrelationships among entities.
This synergy enables the retrieval algorithm to effectively utilize both local and global keywords,
streamlining the search process and improving the relevance of results.
- (i) Query Keyword Extraction. For a given query q, the retrieval algorithm of LightRAG begins
by extracting both local query keywords k[(][l][)] and global query keywords k[(][g][)].
- (ii) Keyword Matching. The algorithm uses an efficient vector database to match local query
keywords with candidate entities and global query keywords with relations linked to global keys.
-----
- (iii) Incorporating High-Order Relatedness. To enhance the query with higher-order relatedness,
LightRAGfurther gathers neighboring nodes within the local subgraphs of the retrieved graph
elements. This process involves the set {vi|vi ∈V ∧ (vi ∈Nv ∨ _vi ∈Ne)}, where Nv and Ne_
represent the one-hop neighboring nodes of the retrieved nodes v and edges e, respectively.
This dual-level retrieval paradigm not only facilitates efficient retrieval of related entities and relations
through keyword matching, but also enhances the comprehensiveness of results by integrating relevant
structural information from the constructed knowledge graph.
3.3 RETRIEVAL-AUGMENTED ANSWER GENERATION
**Utilization of Retrieved Information. Utilizing the retrieved information ψ(q; D[ˆ]), our LightRAG**
employs a general-purpose LLM to generate answers based on the collected data. This data comprises
concatenated values V from relevant entities and relations, produced by the profiling function P(·). It
includes names, descriptions of entities and relations, and excerpts from the original text.
**Context Integration and Answer Generation. By unifying the query with this multi-source text,**
the LLM generates informative answers tailored to the user’s needs, ensuring alignment with the
query’s intent. This approach streamlines the answer generation process by integrating both context
and query into the LLM model, as illustrated in detailed examples (Appendix 7.2).
3.4 COMPLEXITY ANALYSIS OF THE LIGHTRAG FRAMEWORK
In this section, we analyze the complexity of our proposed LightRAG framework, which can be
divided into two main parts. The first part is the graph-based Index phase. During this phase, we use
the large language model (LLM) to extract entities and relationships from each chunk of text. As
a result, the LLM needs to be called [total tokens]chunk size [times. Importantly, there is no additional overhead]
involved in this process, making our approach highly efficient in managing updates to new text.
The second part of the process involves the graph-based retrieval phase. For each query, we first
utilize the large language model (LLM) to generate relevant keywords. Similar to current RetrievalAugmented Generation (RAG) systems Gao et al. (2023; 2022); Chan et al. (2024), our retrieval
mechanism relies on vector-based search. However, instead of retrieving chunks as in conventional
RAG, we concentrate on retrieving entities and relationships. This approach markedly reduces
retrieval overhead compared to the community-based traversal method used in GraphRAG.
#### 4 EVALUATION
We conduct empirical evaluations on benchmark data to assess the effectiveness of the proposed
LightRAG framework by addressing the following research questions: • (RQ1): How does LightRAG
compare to existing RAG baseline methods in terms of generation performance? • (RQ2): How do
dual-level retrieval and graph-based indexing enhance the generation quality of LightRAG? • (RQ3):
What specific advantages does LightRAG demonstrate through case examples in various scenarios? •
**(RQ4): What are the costs associated with LightRAG, as well as its adaptability to data changes?**
4.1 EXPERIMENTAL SETTINGS
**Evaluation Datasets. To conduct a comprehensive analysis of LightRAG, we selected four datasets**
from the UltraDomain benchmark (Qian et al., 2024). The UltraDomain data is sourced from 428
college textbooks and encompasses 18 distinct domains, including agriculture, social sciences, and
humanities. From these, we chose the Agriculture, CS, Legal, and Mix datasets. Each dataset contains
between 600,000 and 5,000,000 tokens, with detailed information provided in Table 4. Below is a
specific introduction to the four domains utilized in our experiments:
- Agriculture: This domain focuses on agricultural practices, covering a range of topics including
beekeeping, hive management, crop production, and disease prevention.
- CS: This domain focuses on computer science and encompasses key areas of data science and
software engineering. It particularly highlights machine learning and big data processing, featuring
content on recommendation systems, classification algorithms, and real-time analytics using Spark.
-----
- Legal: This domain centers on corporate legal practices, addressing corporate restructuring, legal
agreements, regulatory compliance, and governance, with a focus on the legal and financial sectors.
- Mixed: This domain presents a rich variety of literary, biographical, and philosophical texts,
spanning a broad spectrum of disciplines, including cultural, historical, and philosophical studies.
**Question Generation. To evaluate the effectiveness of RAG systems for high-level sensemaking**
tasks, we consolidate all text content from each dataset as context and adopt the generation method
outlined in Edge et al. (2024). Specifically, we instruct an LLM to generate five RAG users, along
with five tasks for each user. Each generated user is accompanied by a textual description detailing
their expertise and traits that motivate their question-raising activities. Each user task is also described,
emphasizing one of the user’s potential intentions when interacting with RAG systems. For each
user-task combination, the LLM generates five questions that require an understanding of the entire
corpus. In total, this process results in 125 questions for each dataset.
**Baselines. LightRAG is compared against the following state-of-the-art methods across all datasets:**
- Naive RAG (Gao et al., 2023): This model serves as a standard baseline in existing RAG systems.
It segments raw texts into chunks and stores them in a vector database using text embeddings. For
queries, Naive RAG generates vectorized representations to directly retrieve text chunks based on
the highest similarity in their representations, ensuring efficient and straightforward matching.
- RQ-RAG (Chan et al., 2024): This approach leverages the LLM to decompose the input query
into multiple sub-queries. These sub-queries are designed to enhance search accuracy by utilizing
explicit techniques such as rewriting, decomposition, and disambiguation.
- HyDE (Gao et al., 2022): This method utilizes the LLM to generate a hypothetical document based
on the input query. This generated document is then employed to retrieve relevant text chunks,
which are subsequently used to formulate the final answer.
- GraphRAG (Edge et al., 2024): This is a graph-enhanced RAG system that utilizes an LLM to
extract entities and relationships from the text, representing them as nodes and edges. It generates
corresponding descriptions for these elements, aggregates nodes into communities, and produces a
community report to capture global information. When handling high-level queries, GraphRAG
retrieves more comprehensive information by traversing these communities.
**[Implementation and Evaluation Details. In our experiments, we utilize the nano vector database](https://github.com/gusye1234/nano-vectordb)**
for vector data management and access. For all LLM-based operations in LightRAG, we default
to using GPT-4o-mini. To ensure consistency, the chunk size is set to 1200 across all datasets.
Additionally, the gleaning parameter is fixed at 1 for both GraphRAG and LightRAG.
Defining ground truth for many RAG queries, particularly those involving complex high-level
semantics, poses significant challenges. To address this, we build on existing work (Edge et al.,
2024) and adopt an LLM-based multi-dimensional comparison method. We employ a robust LLM,
specifically GPT-4o-mini, to rank each baseline against our LightRAG. The evaluation prompt we
used is detailed in Appendix 7.3.4. In total, we utilize four evaluation dimensions, including:
i) Comprehensiveness: How thoroughly does the answer address all aspects and details of the
question? ii) Diversity: How varied and rich is the answer in offering different perspectives and
insights related to the question? iii) Empowerment: How effectively does the answer enable the
reader to understand the topic and make informed judgments? iv) Overall: This dimension assesses
the cumulative performance across the three preceding criteria to identify the best overall answer.
The LLM directly compares two answers for each dimension and selects the superior response for
each criterion. After identifying the winning answer for the three dimensions, the LLM combines the
results to determine the overall better answer. To ensure a fair evaluation and mitigate the potential
bias that could arise from the order in which the answers are presented in the prompt, we alternate the
placement of each answer. We calculate win rates accordingly, ultimately leading to the final results.
4.2 COMPARISON OF LIGHTRAG WITH EXISTING RAG METHODS (RQ1)
We compare LightRAG against each baseline across various evaluation dimensions and datasets. The
results are presented in Table 1. Based on these findings, we draw the following conclusions:
-----
Table 1: Win rates (%) of baselines v.s. LightRAG across four datasets and four evaluation dimensions.
**Agriculture** **CS** **Legal** **Mix**
NaiveRAG **LightRAG** NaiveRAG **LightRAG** NaiveRAG **LightRAG** NaiveRAG **LightRAG**
Comprehensiveness 32.4% 67.6% 38.4% 61.6% 16.4% 83.6% 38.8% 61.2%
Diversity 23.6% 76.4% 38.0% 62.0% 13.6% 86.4% 32.4% 67.6%
Empowerment 32.4% 67.6% 38.8% 61.2% 16.4% 83.6% 42.8% 57.2%
Overall 32.4% 67.6% 38.8% 61.2% 15.2% 84.8% 40.0% 60.0%
RQ-RAG **LightRAG** RQ-RAG **LightRAG** RQ-RAG **LightRAG** RQ-RAG **LightRAG**
Comprehensiveness 31.6% 68.4% 38.8% 61.2% 15.2% 84.8% 39.2% 60.8%
Diversity 29.2% 70.8% 39.2% 60.8% 11.6% 88.4% 30.8% 69.2%
Empowerment 31.6% 68.4% 36.4% 63.6% 15.2% 84.8% 42.4% 57.6%
Overall 32.4% 67.6% 38.0% 62.0% 14.4% 85.6% 40.0% 60.0%
HyDE **LightRAG** HyDE **LightRAG** HyDE **LightRAG** HyDE **LightRAG**
Comprehensiveness 26.0% 74.0% 41.6% 58.4% 26.8% 73.2% 40.4% 59.6%
Diversity 24.0% 76.0% 38.8% 61.2% 20.0% 80.0% 32.4% 67.6%
Empowerment 25.2% 74.8% 40.8% 59.2% 26.0% 74.0% 46.0% 54.0%
Overall 24.8% 75.2% 41.6% 58.4% 26.4% 73.6% 42.4% 57.6%
GraphRAG **LightRAG** GraphRAG **LightRAG** GraphRAG **LightRAG** GraphRAG **LightRAG**
Comprehensiveness 45.6% 54.4% 48.4% 51.6% 48.4% 51.6% 50.4% 49.6%
Diversity 22.8% 77.2% 40.8% 59.2% 26.4% 73.6% 36.0% 64.0%
Empowerment 41.2% 58.8% 45.2% 54.8% 43.6% 56.4% 50.8% 49.2%
Overall 45.2% 54.8% 48.0% 52.0% 47.2% 52.8% 50.4% 49.6%
**The Superiority of Graph-enhanced RAG Systems in Large-Scale Corpora When handling large**
token counts and complex queries that require a thorough understanding of the dataset’s context,
graph-based RAG systems like LightRAG and GraphRAG consistently outperform purely chunkbased retrieval methods such as NaiveRAG, HyDE, and RQRAG. This performance gap becomes
particularly pronounced as the dataset size increases. For instance, in the largest dataset (Legal), the
disparity widens significantly, with baseline methods achieving only about 20% win rates compared
to the dominance of LightRAG. This trend underscores the advantages of graph-enhanced RAG
systems in capturing complex semantic dependencies within large-scale corpora, facilitating a more
comprehensive understanding of knowledge and leading to improved generalization performance.
**Enhancing Response Diversity with LightRAG: Compared to various baselines, LightRAG demon-**
strates a significant advantage in the Diversity metric, particularly within the larger Legal dataset.
Its consistent lead in this area underscores LightRAG’s effectiveness in generating a wider range
of responses, especially in scenarios where diverse content is essential. We attribute this advantage
to LightRAG’s dual-level retrieval paradigm, which facilitates comprehensive information retrieval
from both low-level and high-level dimensions. This approach effectively leverages graph-based text
indexing to consistently capture the full context in response to queries.
**LightRAG’s Superiority over GraphRAG: While both LightRAG and GraphRAG use graph-based**
retrieval mechanisms, LightRAG consistently outperforms GraphRAG, particularly in larger datasets
with complex language contexts. In the Agriculture, CS, and Legal datasets—each containing millions
of tokens—LightRAG shows a clear advantage, significantly surpassing GraphRAG and highlighting
its strength in comprehensive information understanding within diverse environments. Enhanced
**Response Variety: By integrating low-level retrieval of specific entities with high-level retrieval of**
broader topics, LightRAG boosts response diversity. This dual-level mechanism effectively addresses
both detailed and abstract queries, ensuring a thorough grasp of information. Complex Query
**Handling: This approach is especially valuable in scenarios requiring diverse perspectives. By**
accessing both specific details and overarching themes, LightRAG adeptly responds to complex
queries involving interconnected topics, providing contextually relevant answers.
4.3 ABLATION STUDIES (RQ2)
We also conduct ablation studies to evaluate the impact of our dual-level retrieval paradigm and the
effectiveness of our graph-based text indexing in LightRAG. The results are presented in Table 2.
**Effectiveness of Dual-level Retrieval Paradigm. We begin by analyzing the effects of low-level and**
high-level retrieval paradigms. We compare two ablated models—each omitting one module—against
LightRAG across four datasets. Here are our key observations for the different variants:
-----
Table 2: Performance of ablated versions of LightRAG, using NaiveRAG as reference.
**Agriculture** **CS** **Legal** **Mix**
NaiveRAG **LightRAG** NaiveRAG **LightRAG** NaiveRAG **LightRAG** NaiveRAG **LightRAG**
Comprehensiveness 32.4% 67.6% 38.4% 61.6% 16.4% 83.6% 38.8% 61.2%
Diversity 23.6% 76.4% 38.0% 62.0% 13.6% 86.4% 32.4% 67.6%
Empowerment 32.4% 67.6% 38.8% 61.2% 16.4% 83.6% 42.8% 57.2%
Overall 32.4% 67.6% 38.8% 61.2% 15.2% 84.8% 40.0% 60.0%
NaiveRAG **-High** NaiveRAG **-High** NaiveRAG **-High** NaiveRAG **-High**
Comprehensiveness 34.8% 65.2% 42.8% 57.2% 23.6% 76.4% 40.4% 59.6%
Diversity 27.2% 72.8% 36.8% 63.2% 16.8% 83.2% 36.0% 64.0%
Empowerment 36.0% 64.0% 42.4% 57.6% 22.8% 77.2% 47.6% 52.4%
Overall 35.2% 64.8% 44.0% 56.0% 22.0% 78.0% 42.4% 57.6%
NaiveRAG **-Low** NaiveRAG **-Low** NaiveRAG **-Low** NaiveRAG **-Low**
Comprehensiveness 36.0% 64.0% 43.2% 56.8% 19.2% 80.8% 36.0% 64.0%
Diversity 28.0% 72.0% 39.6% 60.4% 13.6% 86.4% 33.2% 66.8%
Empowerment 34.8% 65.2% 42.8% 57.2% 16.4% 83.6% 35.2% 64.8%
Overall 34.8% 65.2% 43.6% 56.4% 18.8% 81.2% 35.2% 64.8%
NaiveRAG **-Origin** NaiveRAG **-Origin** NaiveRAG **-Origin** NaiveRAG **-Origin**
Comprehensiveness 24.8% 75.2% 39.2% 60.8% 16.4% 83.6% 44.4% 55.6%
Diversity 26.4% 73.6% 44.8% 55.2% 14.4% 85.6% 25.6% 74.4%
Empowerment 32.0% 68.0% 43.2% 56.8% 17.2% 82.8% 45.2% 54.8%
Overall 25.6% 74.4% 39.2% 60.8% 15.6% 84.4% 44.4% 55.6%
- Low-level-only Retrieval: The -High variant removes high-order retrieval, leading to a significant
performance decline across nearly all datasets and metrics. This drop is mainly due to its emphasis
on the specific information, which focuses excessively on entities and their immediate neighbors.
While this approach enables deeper exploration of directly related entities, it struggles to gather
information for complex queries that demand comprehensive insights.
- High-level-only Retrieval: The -Low variant prioritizes capturing a broader range of content by
leveraging entity-wise relationships rather than focusing on specific entities. This approach offers
a significant advantage in comprehensiveness, allowing it to gather more extensive and varied
information. However, the trade-off is a reduced depth in examining specific entities, which can
limit its ability to provide highly detailed insights. Consequently, this high-level-only retrieval
method may struggle with tasks that require precise, detailed answers.
- Hybrid Mode: The hybrid mode, or the full version of LightRAG, combines the strengths of
both low-level and high-level retrieval methods. It retrieves a broader set of relationships while
simultaneously conducting an in-depth exploration of specific entities. This dual-level approach
ensures both breadth in the retrieval process and depth in the analysis, providing a comprehensive
view of the data. As a result, LightRAG achieves balanced performance across multiple dimensions.
**Semantic Graph Excels in RAG. We eliminated the use of original text in our retrieval process.**
Surprisingly, the resulting variant, -Origin, does not exhibit significant performance declines across
all four datasets. In some cases, this variant even shows improvements (e.g. in Agriculture and Mix).
We attribute this phenomenon to the effective extraction of key information during the graph-based
indexing process, which provides sufficient context for answering queries. Additionally, the original
text often contains irrelevant information that can introduce noise in the response.
4.4 CASE STUDY (RQ3)
To provide a clear comparison between baseline methods and our LightRAG, we present specific
case examples in Table 3, which includes responses to a machine learning question from both
the competitive baseline, GraphRAG, and our LightRAG framework. In this instance, LightRAG
outperforms in all evaluation dimensions assessed by the LLM judge, including comprehensiveness,
diversity, empowerment, and overall quality. Our key observations are as follows:
**i) Comprehensiveness. Notably, LightRAG covers a broader range of machine learning metrics,**
showcasing its comprehensiveness and ability to effectively discover relevant information. This
highlights the strength of our graph-based indexing paradigm, which excels in precise entity and
relation extraction as well as LLM profiling. ii) Both Diversity and Empowerment. Furthermore,
LightRAG not only offers a more diverse array of information but also delivers more empowering
content. This success is due to LightRAG’s hierarchical retrieval paradigm, which combines in-depth
-----
Table 3: Case Study: Comparison Between LightRAG and the Baseline Method GraphRAG.
**Query** Which methods can normalize the feature values for effective machine learning?
**GraphRAG** **Key Metrics for Evaluating Movie Recommendation Systems: Evaluating the effectiveness of movie**
recommendation systems is crucial for ensuring user satisfaction and improving the overall performance of
these systems. Several metrics are widely recognized for their ability to provide insights into how well a
recommendation system performs. Below are the most informative metrics: 1. Precision [...] 2. Recall [...]
3. F1 Score [...] 4. Mean Average Precision (MAP) [...] 5. Root Mean Squared Error (RMSE) [...] 6.
**User Satisfaction Metrics [...] Conclusion: In summary, metrics such as precision, recall, F1 score, MAP,**
RMSE, and user satisfaction are critical for evaluating the performance of movie recommendation systems.
Each metric provides unique insights that can help developers and researchers refine their algorithms and
enhance user experiences. By focusing on these metrics, stakeholders can ensure that their recommendation
systems effectively meet user preferences and expectations.
**LightRAG** **Key Metrics for Evaluating Movie Recommendation Systems: Evaluating movie recommendation systems**
involves multiple metrics that assess the effectiveness and accuracy of the recommendations provided to
users. Below are some of the most informative metrics commonly used in this context, derived from the
data and concepts surrounding recommendation systems. 1. Mean Average Precision at K (MAPK) [...] 2.
**Precision and Recall [...] 3. Root Mean Squared Error (RMSE) and Mean Squared Error (MSE) [...]**
4. Area Under the Curve (AUC) [...] 5. F-Measure [...] 6. User Engagement Metrics [...] Conclusion: A
comprehensive evaluation of movie recommendation systems employs a range of metrics that capture both
the accuracy of the recommendations and user satisfaction. By effectively leveraging these metrics, system
designers can optimize recommendation engines to deliver more personalized and relevant experiences.
**LLM Decision** **Comprehensiveness: "Winner": "Answer 2 (LightRAG)",**
"Explanation": "Answer 2 provides a broader array of metrics including MAPK, AUC, and user engagement
metrics alongside precision, recall, RMSE, and F-Measure. This gives a fuller view of how recommendation
systems can be evaluated. Answer 1, while detailed, is more focused on traditional metrics and lacks the
inclusion of engagement metrics which are crucial for understanding user interaction."
**Diversity: "Winner": "Answer 2 (LightRAG)",**
"Explanation": "Answer 2 not only covers a wide variety of metrics but also includes nuanced explanations
of how some metrics interrelate and differ from one another, like the inclusion of both RMSE and MSE, as
well as the details behind AUC. In contrast, Answer 1 sticks primarily to standard metrics without much
exploration of potential nuances."
**Empowerment: "Winner": "Answer 2 (LightRAG)",**
"Explanation": "Answer 2 empowers the reader more effectively by detailing how each metric functions and
its importance in evaluating recommendation systems. By providing context such as the trade-offs between
precision and recall and emphasizing user engagement metrics, it enables readers to make more informed
judgments and understand the implications of different metrics. Answer 1 is more straightforward but lacks
the depth of insight regarding why these metrics matter."
**Overall Winner: "Winner": "Answer 2 (LightRAG)",**
"Explanation": "While Answer 1 is more direct and systematic, Answer 2 excels in comprehensiveness,
diversity, and empowerment. It provides a richer exploration of the topic, including insights into user
engagement and nuanced differences between metrics. This depth and breadth make it more informative for
readers seeking to thoroughly understand the evaluation of movie recommendation systems."
explorations of related entities through low-level retrieval to enhance empowerment with broader
explorations via high-level retrieval to improve answer diversity. Together, these approaches capture a
comprehensive global perspective of the knowledge domain, contributing to better RAG performance.
4.5 MODEL COST AND ADAPTABILITY ANALYSIS (RQ4)
We compare the cost of our LightRAG with that of the Figure 2: Comparison of Cost in Terms of Totop-performing baseline, GraphRAG, from two key kens and API Calls for GraphRAG and Lighperspectives. First, we examine the number of tokens tRAG on the Legal Dataset.
and API calls during the indexing and retrieval pro- Phase Retrieval Phase Incremental Text Update
cesses. Second, we analyze these metrics in relation Model GraphRAG Ours GraphRAG Ours
The results of this evaluation on the legal dataset areto handling data changes in dynamic environments.presented in Table 2. In this context, Textract repre- TokensCallsAPI 610610 ×C×max1 1,000,000 < 100 1 1[1],[,]399[399] ×+[ ×]T 2 +[ 2]extract[ ×] C[ 5][,]extract[000] CTextractextract
sents the token overhead for entity and relationship extraction, Cmax denotes the maximum number
of tokens allowed per API call, and Cextract indicates the number of API calls required for extraction.
In the retrieval phase, GraphRAG generates 1,399 communities, with 610 level-2 communities
actively utilized for retrieval in this experiment. Each community report averages 1,000 tokens,
resulting in a total token consumption of 610,000 tokens (610 communities × 1,000 tokens per
community). Additionally, GraphRAG’s requirement to traverse each community individually leads
to hundreds of API calls, significantly increasing retrieval overhead. In contrast, LightRAG optimizes
this process by using fewer than 100 tokens for keyword generation and retrieval, requiring only a
single API call for the entire process. This efficiency is achieved through our retrieval mechanism,
|Query|Which methods can normalize the feature values for effective machine learning?|
|---|---|
|GraphRAG|Key Metrics for Evaluating Movie Recommendation Systems: Evaluating the effectiveness of movie recommendation systems is crucial for ensuring user satisfaction and improving the overall performance of these systems. Several metrics are widely recognized for their ability to provide insights into how well a recommendation system performs. Below are the most informative metrics: 1. Precision [...] 2. Recall [...] 3. F1 Score [...] 4. Mean Average Precision (MAP) [...] 5. Root Mean Squared Error (RMSE) [...] 6. User Satisfaction Metrics [...] Conclusion: In summary, metrics such as precision, recall, F1 score, MAP, RMSE, and user satisfaction are critical for evaluating the performance of movie recommendation systems. Each metric provides unique insights that can help developers and researchers refine their algorithms and enhance user experiences. By focusing on these metrics, stakeholders can ensure that their recommendation systems effectively meet user preferences and expectations.|
|LightRAG|Key Metrics for Evaluating Movie Recommendation Systems: Evaluating movie recommendation systems involves multiple metrics that assess the effectiveness and accuracy of the recommendations provided to users. Below are some of the most informative metrics commonly used in this context, derived from the data and concepts surrounding recommendation systems. 1. Mean Average Precision at K (MAPK) [...] 2. Precision and Recall [...] 3. Root Mean Squared Error (RMSE) and Mean Squared Error (MSE) [...] 4. Area Under the Curve (AUC) [...] 5. F-Measure [...] 6. User Engagement Metrics [...] Conclusion: A comprehensive evaluation of movie recommendation systems employs a range of metrics that capture both the accuracy of the recommendations and user satisfaction. By effectively leveraging these metrics, system designers can optimize recommendation engines to deliver more personalized and relevant experiences.|
|LLM Decision|Comprehensiveness: "Winner": "Answer 2 (LightRAG)", "Explanation": "Answer 2 provides a broader array of metrics including MAPK, AUC, and user engagement metrics alongside precision, recall, RMSE, and F-Measure. This gives a fuller view of how recommendation systems can be evaluated. Answer 1, while detailed, is more focused on traditional metrics and lacks the inclusion of engagement metrics which are crucial for understanding user interaction." Diversity: "Winner": "Answer 2 (LightRAG)", "Explanation": "Answer 2 not only covers a wide variety of metrics but also includes nuanced explanations of how some metrics interrelate and differ from one another, like the inclusion of both RMSE and MSE, as well as the details behind AUC. In contrast, Answer 1 sticks primarily to standard metrics without much exploration of potential nuances." Empowerment: "Winner": "Answer 2 (LightRAG)", "Explanation": "Answer 2 empowers the reader more effectively by detailing how each metric functions and its importance in evaluating recommendation systems. By providing context such as the trade-offs between precision and recall and emphasizing user engagement metrics, it enables readers to make more informed judgments and understand the implications of different metrics. Answer 1 is more straightforward but lacks the depth of insight regarding why these metrics matter." Overall Winner: "Winner": "Answer 2 (LightRAG)", "Explanation": "While Answer 1 is more direct and systematic, Answer 2 excels in comprehensiveness, diversity, and empowerment. It provides a richer exploration of the topic, including insights into user engagement and nuanced differences between metrics. This depth and breadth make it more informative for readers seeking to thoroughly understand the evaluation of movie recommendation systems."|
|Phase|Retrieval Phase|Col3|Incremental Text Update|Col5|
|---|---|---|---|---|
|Model|GraphRAG|Ours|GraphRAG|Ours|
|Tokens|610 × 1,000|< 100|1,399 × 2 × 5,000 +Textract|Textract|
|API Calls|610×1,000 Cmax|1|1,399 × 2 + Cextract|Cextract|
-----
which seamlessly integrates graph structures and vectorized representations for information retrieval,
thereby eliminating the need to process large volumes of information upfront.
In the incremental data update phase, designed to address changes in dynamic real-world scenarios,
both models exhibit similar overhead for entity and relationship extraction. However, GraphRAG
shows significant inefficiency in managing newly added data. When a new dataset of the same
size as the legal dataset is introduced, GraphRAG must dismantle its existing community structure
to incorporate new entities and relationships, followed by complete regeneration. This process
incurs a substantial token cost of approximately 5,000 tokens per community report. Given 1,399
communities, GraphRAG would require around 1,399 × 2 × 5,000 tokens to reconstruct both the
original and new community reports—an exorbitant expense that underscores its inefficiency. In
contrast, LightRAG seamlessly integrates newly extracted entities and relationships into the existing
graph without the need for full reconstruction. This approach results in significantly lower overhead
during incremental updates, demonstrating its superior efficiency and cost-effectiveness.
#### 5 RELATED WORK
5.1 RETRIEVAL-AUGMENTED GENERATION WITH LLMS
Retrieval-Augmented Generation (RAG) systems enhance LLM inputs by retrieving relevant information from external sources, grounding responses in factual, domain-specific knowledge Ram et al.
(2023); Fan et al. (2024). Current RAG approaches Gao et al. (2022; 2023); Chan et al. (2024); Yu
et al. (2024) typically embed queries in a vector space to find the nearest context vectors. However,
many of these methods rely on fragmented text chunks and only retrieve the top-k contexts, limiting
their ability to capture comprehensive global information needed for effective responses.
Although recent studies Edge et al. (2024) have explored using graph structures for knowledge
representation, two key limitations persist. First, these approaches often lack the capability for
dynamic updates and expansions of the knowledge graph, making it difficult to incorporate new
information effectively. In contrast, our proposed model, LightRAG, addresses these challenges
by enabling the RAG system to quickly adapt to new information, ensuring the model’s timeliness
and accuracy. Additionally, existing methods often rely on brute-force searches for each generated
community, which are inefficient for large-scale queries. Our LightRAG framework overcomes this
limitation by facilitating rapid retrieval of relevant information from the graph through our proposed
dual-level retrieval paradigm, significantly enhancing both retrieval efficiency and response speed.
5.2 LARGE LANGUAGE MODEL FOR GRAPHS
Graphs are a powerful framework for representing complex relationships and find applications
in numerous fields. As Large Language Models (LLMs) continue to evolve, researchers have
increasingly focused on enhancing their capability to interpret graph-structured data. This body of
work can be divided into three primary categories: i) GNNs as Prefix where Graph Neural Networks
(GNNs) are utilized as the initial processing layer for graph data, generating structure-aware tokens
that LLMs can use during inference. Notable examples include GraphGPT Tang et al. (2024) and
LLaGA Chen et al. (2024). ii) LLMs as Prefix involves LLMs processing graph data enriched with
textual information to produce node embeddings or labels, ultimately refining the training process
for GNNs, as demonstrated in systems like GALM Xie et al. (2023) and OFA Liu et al. (2024). iii)
**LLMs-Graphs Integration focuses on achieving a seamless interaction between LLMs and graph**
data, employing techniques such as fusion training and GNN alignment, and developing LLM-based
agents capable of engaging with graph information directly Li et al. (2023); Brannon et al. (2023).
#### 6 CONCLUSION
This work introduces an advancement in Retrieval-Augmented Generation (RAG) through the
integration of a graph-based indexing approach that enhances both efficiency and comprehension
in information retrieval. LightRAG utilizes a comprehensive knowledge graph to facilitate rapid
and relevant document retrieval, enabling a deeper understanding of complex queries. Its dual-level
retrieval paradigm allows for the extraction of both specific and abstract information, catering to
diverse user needs. Furthermore, LightRAG’s seamless incremental update capability ensures that
the system remains current and responsive to new information, thereby maintaining its effectiveness
over time. Overall, LightRAG excels in both efficiency and effectiveness, significantly improving the
speed and quality of information retrieval and generation while reducing costs for LLM inference.
-----
#### REFERENCES
William Brannon, Suyash Fulay, Hang Jiang, Wonjune Kang, Brandon Roy, Jad Kabbara, and Deb
Roy. Congrat: Self-supervised contrastive pretraining for joint graph and text embeddings. arXiv
_preprint arXiv:2305.14321, 2023._
Chi-Min Chan, Chunpu Xu, Ruibin Yuan, Hongyin Luo, Wei Xue, Yike Guo, and Jie Fu. Rq-rag:
Learning to refine queries for retrieval augmented generation. arXiv preprint arXiv:2404.00610,
2024.
Runjin Chen, Tong Zhao, AJAY KUMAR JAISWAL, Neil Shah, and Zhangyang Wang. Llaga: Large
language and graph assistant. In International Conference on Machine Learning (ICML), 2024.
Darren Edge, Ha Trinh, Newman Cheng, Joshua Bradley, Alex Chao, Apurva Mody, Steven Truitt,
and Jonathan Larson. From local to global: A graph rag approach to query-focused summarization.
_arXiv preprint arXiv:2404.16130, 2024._
Shahul Es, Jithin James, Luis Espinosa Anke, and Steven Schockaert. Ragas: Automated evaluation
of retrieval augmented generation. In International Conference of the European Chapter of the
_Association for Computational Linguistics (EACL), pp. 150–158, 2024._
Wenqi Fan, Yujuan Ding, Liangbo Ning, Shijie Wang, Hengyun Li, Dawei Yin, Tat-Seng Chua, and
Qing Li. A survey on rag meeting llms: Towards retrieval-augmented large language models. In
_International Conference on Knowledge Discovery and Data Mining (KDD), pp. 6491–6501, 2024._
Luyu Gao, Xueguang Ma, Jimmy Lin, and Jamie Callan. Precise zero-shot dense retrieval without
relevance labels. arXiv preprint arXiv:2212.10496, 2022.
Yunfan Gao, Yun Xiong, Xinyu Gao, Kangxiang Jia, Jinliu Pan, Yuxi Bi, Yi Dai, Jiawei Sun, and
Haofen Wang. Retrieval-augmented generation for large language models: A survey. arXiv
_preprint arXiv:2312.10997, 2023._
Yichuan Li, Kaize Ding, and Kyumin Lee. Grenade: Graph-centric language model for selfsupervised representation learning on text-attributed graphs. In International Conference on
_Empirical Methods in Natural Language Processing (EMNLP), pp. 2745–2757, 2023._
Hao Liu, Jiarui Feng, Lecheng Kong, Ningyue Liang, Dacheng Tao, Yixin Chen, and Muhan Zhang.
One for all: Towards training one graph model for all classification tasks. In International
_Conference on Learning Representations (ICLR), 2024._
Yuanjie Lyu, Zhiyu Li, Simin Niu, Feiyu Xiong, Bo Tang, Wenjin Wang, Hao Wu, Huanyong
Liu, Tong Xu, and Enhong Chen. Crud-rag: A comprehensive chinese benchmark for retrievalaugmented generation of large language models. arXiv preprint arXiv:2401.17043, 2024.
Hongjin Qian, Peitian Zhang, Zheng Liu, Kelong Mao, and Zhicheng Dou. Memorag: Moving
[towards next-gen rag via memory-inspired knowledge discovery, 2024. URL https://arxiv.](https://arxiv.org/abs/2409.05591)
[org/abs/2409.05591.](https://arxiv.org/abs/2409.05591)
Ori Ram, Yoav Levine, Itay Dalmedigos, Dor Muhlgay, Amnon Shashua, Kevin Leyton-Brown, and
Yoav Shoham. In-context retrieval-augmented language models. Transactions of the Association
_for Computational Linguistics (TACL), 11:1316–1331, 2023._
Ladislav Rampášek, Michael Galkin, Vijay Prakash Dwivedi, Anh Tuan Luu, Guy Wolf, and Dominique Beaini. Recipe for a general, powerful, scalable graph transformer. _International_
_Conference on Neural Information Processing Systems (NeurIPS), 35:14501–14515, 2022._
Alireza Salemi and Hamed Zamani. Evaluating retrieval quality in retrieval-augmented generation.
In ACM International Conference on Research and Development in Information Retrieval (SIGIR),
pp. 2395–2400, 2024.
Viju Sudhi, Sinchana Ramakanth Bhat, Max Rudat, and Roman Teucher. Rag-ex: A generic
framework for explaining retrieval augmented generation. In ACM International Conference on
_Research and Development in Information Retrieval (SIGIR), pp. 2776–2780, 2024._
-----
Jiabin Tang, Yuhao Yang, Wei Wei, Lei Shi, Lixin Su, Suqi Cheng, Dawei Yin, and Chao Huang.
Graphgpt: Graph instruction tuning for large language models. In ACM International Conference
_on Research and Development in Information Retrieval (SIGIR), pp. 491–500, 2024._
Shangqing Tu, Yuanchun Wang, Jifan Yu, Yuyang Xie, Yaran Shi, Xiaozhi Wang, Jing Zhang, Lei
Hou, and Juanzi Li. R-eval: A unified toolkit for evaluating domain knowledge of retrieval
augmented large language models. In International Conference on Knowledge Discovery and Data
_Mining (KDD), pp. 5813–5824, 2024._
Han Xie, Da Zheng, Jun Ma, Houyu Zhang, Vassilis N Ioannidis, Xiang Song, Qing Ping, Sheng
Wang, Carl Yang, Yi Xu, et al. Graph-aware language model pre-training on a large graph corpus
can help multiple graph applications. In International Conference on Knowledge Discovery and
_Data Mining (KDD), pp. 5270–5281, 2023._
Yue Yu, Wei Ping, Zihan Liu, Boxin Wang, Jiaxuan You, Chao Zhang, Mohammad Shoeybi, and
Bryan Catanzaro. Rankrag: Unifying context ranking with retrieval-augmented generation in llms.
_arXiv preprint arXiv:2407.02485, 2024._
Penghao Zhao, Hailin Zhang, Qinhan Yu, Zhengren Wang, Yunteng Geng, Fangcheng Fu, Ling Yang,
Wentao Zhang, and Bin Cui. Retrieval-augmented generation for ai-generated content: A survey.
_arXiv preprint arXiv:2402.19473, 2024._
-----
#### 7 APPENDIX
In this section, we elaborate on the methodologies and experimental settings used in the LightRAG
framework. It describes the specific steps for extracting entities and relationships from documents,
detailing how large language models (LLMs) are utilized for this purpose. The section also specifies
the prompt templates and configurations used in LLM operations, ensuring clarity in the experimental
setup. Additionally, it outlines the evaluation criteria and dimensions used to assess the performance
of LightRAG against baselines from various dimensions.
7.1 EXPERIMENTAL DATA DETAILS
Table 4: Statistical information of the datasets.
**Statistics** **Agriculture** **CS** **Legal** **Mix**
Total Documents 12 10 94 61
Total Tokens 2,017,886 2,306,535 5,081,069 619,009
Table 4 presents statistical information for four datasets: Agriculture, CS, Legal, and Mix. The
Agriculture dataset consists of 12 documents totaling 2,017,886 tokens, while the CS dataset contains
10 documents with 2,306,535 tokens. The Legal dataset is the largest, comprising 94 documents and
5,081,069 tokens. Lastly, the Mix dataset includes 61 documents with a total of 619,009 tokens.
7.2 CASE EXAMPLE OF RETRIEVAL-AUGMENTED GENERATION IN LIGHTRAG.
Figure 3: A retrieval and generation example.
In Figure 3, we illustrate the retrieve-and-generate process. When presented with the query, “What
metrics are most informative for evaluating movie recommendation systems?”, the LLM first extracts
both low-level and high-level keywords. These keywords guide the dual-level retrieval process on the
-----
generated knowledge graph, targeting relevant entities and relationships. The retrieved information
is organized into three components: entities, relationships, and corresponding text chunks. This
structured data is then fed into the LLM, enabling it to generate a comprehensive answer to the query.
7.3 OVERVIEW OF THE PROMPTS USED IN LIGHTRAG
7.3.1 PROMPTS FOR GRAPH GENERATION
Figure 4: Prompts for Graph Generation
The graph construction prompt outlined in Figure 4 is designed to extract and structure entityrelationship information from a text document based on specified entity types. The process begins
by identifying entities and categorizing them into types such as organization, person, location, and
event. It then provides detailed descriptions of their attributes and activities. Next, the prompt
identifies relationships between these entities, offering explanations, assigning strength scores, and
summarizing the relationships using high-level keywords.
7.3.2 PROMPTS FOR QUERY GENERATION
Figure 5: Prompts for Query Generation
In Figure 5, the query generation prompt outlines a framework for identifying potential user roles
(e.g., data scientist, finance analyst, and product manager) and their objectives for generating queries
based on a specified dataset description. The prompt explains how to define five distinct users
who would benefit from interacting with the dataset. For each user, it specifies five key tasks they
would perform while working with the dataset. Additionally, for each (user, task) combination, five
high-level questions are posed to ensure a thorough understanding of the dataset.
-----
7.3.3 PROMPTS FOR KEYWORD EXTRACTION
Figure 6: Prompts for Keyword Extraction
In Figure 6, the prompt describes a method for extracting keywords from a user’s query, distinguishing
between high-level and low-level keywords. High-level keywords represent broad concepts or
themes, while low-level keywords focus on specific entities and details. The extracted keywords are
returned in JSON format, organized into two fields: “high_level_keywords” for overarching ideas
and “low_level_keywords” for specific details.
7.3.4 PROMPTS FOR RAG EVALUATION
Figure 7: Prompts for RAG Evaluation
The evaluation prompt is illustrated in Figure 7. It introduces a comprehensive evaluation framework
for comparing two answers to the same question based on three key criteria: Comprehensiveness,
Diversity, and Empowerment. Its purpose is to guide the LLM through the process of selecting the
better answer for each criterion, followed by an overall assessment. For each of the three criteria, the
-----
Table 5: Case Study: Comparison Between LightRAG and the Baseline NaiveRAG.
**Query** How do indigenous perspectives on ownership and collaboration influence corporate mergers in countries
like Canada and Australia?
**NaiveRAG** Indigenous perspectives on ownership and collaboration play a significant role in shaping corporate mergers
in countries like Canada and Australia, particularly where resources and land rights intersect with Indigenous sovereignty and cultural values. These perspectives emphasize community engagement, respect for
traditional land use, and a collaborative approach to resource management, influencing corporate strategies
and outcomes.
...
Both nations illustrate that prioritizing Indigenous perspectives leads to more sustainable and equitable
outcomes. Acknowledging Indigenous rights and fostering collaborative relationships can reduce conflicts
and enhance corporate reputations. Ultimately, integrating Indigenous perspectives is a pathway to better
business practices and community relations.
**LightRAG** **Indigenous Perspectives on Ownership and Collaboration in Corporate Mergers**
Indigenous perspectives significantly influence corporate mergers in countries like Canada and Australia.
These views emphasize communal rights to land and resources, spiritual connections to the environment, and
the importance of collaboration over competition.
**Cultural Significance of Land Ownership [...]**
**The Role of Collaboration [...]**
**Legal and Regulatory Frameworks [...]**
**Conclusion**
Indigenous perspectives shape corporate mergers by emphasizing communal ownership and collaborative
partnerships. Corporations must adopt respectful, inclusive practices that honor Indigenous cultural values
and rights, paving the way for sustainable, mutually beneficial outcomes.
**LLM Decision** **Comprehensiveness: "Winner": "Answer 2 (LightRAG)",**
"Explanation": "Answer 2 provides a thorough exploration of the indigenous perspectives by discussing
cultural significance, collaboration, and legal frameworks, including specific examples and detailed insights.
In contrast, Answer 1, while informative, lacks the same depth in analyzing the various dimensions of the
indigenous perspectives on ownership and collaboration."
**Diversity: "Winner": "Answer 2 (LightRAG)",**
"Explanation": "Answer 2 presents a wider array of perspectives, focusing on the communal aspect of
land ownership, spiritual connections, and practical examples of collaboration. It covers more ground by
contrasting Indigenous views with Western notions, whereas Answer 1 primarily focuses on corporate
strategies and outcomes with limited perspective."
**Empowerment: "Winner": "Answer 2 (LightRAG)",**
"Explanation": "Answer 2 equips the reader with nuanced understanding and actionable insights by
highlighting the importance of collaboration and the legal frameworks involved. It empowers corporations
by illustrating how adopting an inclusive approach can facilitate better outcomes. Answer 1, while practical,
does not emphasize the moral or ethical implications as strongly."
**Overall Winner: "Winner": "Answer 2 (LightRAG)",**
"Explanation": "Answer 2 excels overall due to its comprehensive exploration, diversity of perspectives,
and empowerment of the reader with actionable insights about indigenous perspectives and collaboration
in corporate mergers. Although Answer 1 is more direct, the depth and breadth of Answer 2 make it the
stronger response."
LLM must identify which answer performs better and provide a rationale for its choice. Ultimately,
an overall winner is determined based on performance across all three dimensions, accompanied by a
detailed summary that justifies the decision. The evaluation is structured in JSON format, ensuring
clarity and consistency, and facilitating a systematic comparison between the two answers.
7.4 CASE STUDY: COMPARISON BETWEEN LIGHTRAG AND THE BASELINE NAIVERAG.
To further illustrate LightRAG’s superiority over baseline models in terms of comprehensiveness,
empowerment, and diversity, we present a case study comparing LightRAG and NaiveRAG in
Table 5. This study addresses a question regarding indigenous perspectives in the context of corporate
mergers. Notably, LightRAG offers a more in-depth exploration of key themes related to indigenous
perspectives, such as cultural significance, collaboration, and legal frameworks, supported by specific
and illustrative examples. In contrast, while NaiveRAG provides informative responses, it lacks
the depth needed to thoroughly examine the various dimensions of indigenous ownership and
collaboration. The dual-level retrieval process employed by LightRAG enables a more comprehensive
investigation of specific entities and their interrelationships, facilitating extensive searches that
effectively capture overarching themes and complexities within the topic.
|Query|How do indigenous perspectives on ownership and collaboration influence corporate mergers in countries like Canada and Australia?|
|---|---|
|NaiveRAG|Indigenous perspectives on ownership and collaboration play a significant role in shaping corporate mergers in countries like Canada and Australia, particularly where resources and land rights intersect with Indige- nous sovereignty and cultural values. These perspectives emphasize community engagement, respect for traditional land use, and a collaborative approach to resource management, influencing corporate strategies and outcomes. ... Both nations illustrate that prioritizing Indigenous perspectives leads to more sustainable and equitable outcomes. Acknowledging Indigenous rights and fostering collaborative relationships can reduce conflicts and enhance corporate reputations. Ultimately, integrating Indigenous perspectives is a pathway to better business practices and community relations.|
|LightRAG|Indigenous Perspectives on Ownership and Collaboration in Corporate Mergers Indigenous perspectives significantly influence corporate mergers in countries like Canada and Australia. These views emphasize communal rights to land and resources, spiritual connections to the environment, and the importance of collaboration over competition. Cultural Signifciance of Land Ownership [...] The Role of Collaboration [...] Legal and Regulatory Frameworks [...] Conclusion Indigenous perspectives shape corporate mergers by emphasizing communal ownership and collaborative partnerships. Corporations must adopt respectful, inclusive practices that honor Indigenous cultural values and rights, paving the way for sustainable, mutually beneficial outcomes.|
|LLM Decision|Comprehensiveness: "Winner": "Answer 2 (LightRAG)", "Explanation": "Answer 2 provides a thorough exploration of the indigenous perspectives by discussing cultural significance, collaboration, and legal frameworks, including specific examples and detailed insights. In contrast, Answer 1, while informative, lacks the same depth in analyzing the various dimensions of the indigenous perspectives on ownership and collaboration." Diversity: "Winner": "Answer 2 (LightRAG)", "Explanation": "Answer 2 presents a wider array of perspectives, focusing on the communal aspect of land ownership, spiritual connections, and practical examples of collaboration. It covers more ground by contrasting Indigenous views with Western notions, whereas Answer 1 primarily focuses on corporate strategies and outcomes with limited perspective." Empowerment: "Winner": "Answer 2 (LightRAG)", "Explanation": "Answer 2 equips the reader with nuanced understanding and actionable insights by highlighting the importance of collaboration and the legal frameworks involved. It empowers corporations by illustrating how adopting an inclusive approach can facilitate better outcomes. Answer 1, while practical, does not emphasize the moral or ethical implications as strongly." Overall Winner: "Winner": "Answer 2 (LightRAG)", "Explanation": "Answer 2 excels overall due to its comprehensive exploration, diversity of perspectives, and empowerment of the reader with actionable insights about indigenous perspectives and collaboration in corporate mergers. Although Answer 1 is more direct, the depth and breadth of Answer 2 make it the stronger response."|
-----
|
{
"id": "2410.05779",
"submitter": "Zirui Guo",
"authors": "Zirui Guo, Lianghao Xia, Yanhua Yu, Tu Ao, Chao Huang",
"title": "LightRAG: Simple and Fast Retrieval-Augmented Generation",
"comments": null,
"journal-ref": null,
"doi": null,
"report-no": null,
"categories": "cs.IR cs.AI",
"license": "http://creativecommons.org/licenses/by-nc-sa/4.0/",
"abstract": " Retrieval-Augmented Generation (RAG) systems enhance large language models\n(LLMs) by integrating external knowledge sources, enabling more accurate and\ncontextually relevant responses tailored to user needs. However, existing RAG\nsystems have significant limitations, including reliance on flat data\nrepresentations and inadequate contextual awareness, which can lead to\nfragmented answers that fail to capture complex inter-dependencies. To address\nthese challenges, we propose LightRAG, which incorporates graph structures into\ntext indexing and retrieval processes. This innovative framework employs a\ndual-level retrieval system that enhances comprehensive information retrieval\nfrom both low-level and high-level knowledge discovery. Additionally, the\nintegration of graph structures with vector representations facilitates\nefficient retrieval of related entities and their relationships, significantly\nimproving response times while maintaining contextual relevance. This\ncapability is further enhanced by an incremental update algorithm that ensures\nthe timely integration of new data, allowing the system to remain effective and\nresponsive in rapidly changing data environments. Extensive experimental\nvalidation demonstrates considerable improvements in retrieval accuracy and\nefficiency compared to existing approaches. We have made our LightRAG\nopen-source and available at the link: https://github.com/HKUDS/LightRAG.\n",
"versions": {
"version": [
"v1",
"v2"
],
"created": [
"Tue, 8 Oct 2024 08:00:12 GMT",
"Thu, 7 Nov 2024 10:44:59 GMT"
]
},
"update_date": "2024-11-08",
"authors_parsed": [
[
"Guo",
"Zirui",
""
],
[
"Xia",
"Lianghao",
""
],
[
"Yu",
"Yanhua",
""
],
[
"Ao",
"Tu",
""
],
[
"Huang",
"Chao",
""
]
]
}
|
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.