Datasets:

AdithyaSK
/

RAG_Eval

Large language models (LLMs) inevitably exhibit hallucinations since the accuracy of generated texts cannot be secured solely by the parametric knowledge they encapsulate. Although retrieval-augmented generation (RAG) is a practicable complement to LLMs, it relies heavily on the relevance of retrieved documents, raising concerns about how the model behaves if retrieval goes wrong. To this end, we propose the Corrective Retrieval Augmented Generation (CRAG) to improve the robustness of generation. Specifically, a lightweight retrieval evaluator is designed to assess the overall quality of retrieved documents for a query, returning a confidence degree based on which different knowledge retrieval actions can be triggered. Since retrieval from static and limited corpora can only return sub-optimal documents, large-scale web searches are utilized as an extension for augmenting the retrieval results. Besides, a decompose-then-recompose algorithm is designed for retrieved documents to selectively focus on key information and filter out irrelevant information in them. CRAG is plug-and-play and can be seamlessly coupled with various RAG-based approaches. Experiments on four datasets covering short- and long-form generation tasks show that CRAG can significantly improve the performance of RAG-based approaches.1

1 Introduction

Large language models (LLMs) have attracted increasing attention and exhibited impressive abilities to understand instructions and generate fluent language texts (Brown et al., 2020; Ouyang et al., 2022; Touvron et al., 2023a). Nevertheless, LLMs inevitably manifest hallucinations (Ji et al., 2023) due to their struggle with factual errors (Mallen et al., 2023; Min et al., 2023) and inability to secure the accuracy of generated texts solely by the parametric knowledge they encapsulate (Zhang et al., 2023b; Muhlgay et al., 2023).

Equal contribution.

1The code is available at github.com/HuskyInSalt/CRAG

Retriever

Accurate Documents	Inaccurate Documents
Henry Feilden (Conservative politician): Henry Master Feilden was an Conservative Party politician…	Batman (1989 film): of the murder of Bruce Wayne's parents. When Hamm's script was rewritten, …
Politician.✓	Hamm.✗

GeneratorRobo

Figure 1: The examples show that a low-quality retriever is prone to introducing a substantial amount of irrelevant information, impeding the generators from acquiring accurate knowledge and potentially misleading them.

Prior research has introduced retrieval techniques to incorporate relevant knowledge and augment generation, as exemplified by retrieval-augmented generation (RAG) (Lewis et al., 2020). In this framework, the input to models is augmented by prepending relevant documents that are retrieved from an external knowledge corpus (Guu et al., 2020). While RAG serves as a practicable complement to LLMs, its effectiveness is contingent upon the relevance and accuracy of the retrieved documents (Li et al., 2022; Tan et al., 2022). The heavy reliance of generation on the retrieved knowledge raises significant concerns about the model’s behavior and performance in scenarios where retrieval may fail or return inaccurate results (Shi et al., 2023). As Figure 1 shows that a low-quality retriever is prone to introducing a substantial amount of irrelevant information.

Current Page Title

impeding the models from acquiring accurate knowledge and potentially misleading them, resulting in issues such as hallucinations (Zhang et al., 2023b). However, most conventional RAG approaches indiscriminately incorporate the retrieved documents, regardless of whether these documents are relevant or not (Rony et al., 2022). Furthermore, current methods mostly treat complete documents as reference knowledge both during retrieval and utilization. But a considerable portion of the text within these retrieved documents is often non-essential for generation, which should not have been equally referred to and involved in RAG.

On account of the above issues, this paper particularly studies the scenarios where the retriever returns inaccurate results. A method named Corrective Retrieval-Augmented Generation (CRAG) is proposed to self-correct the results of retriever and improve the utilization of documents for augmenting generation. A lightweight retrieval evaluator is designed to assess the overall quality of retrieved documents for a query. This serves as a crucial component in RAG, contributing to informative generation by reviewing and evaluating the relevance and reliability of the retrieved documents. A confidence degree is quantified based on which different knowledge retrieval actions of {Correct, Incorrect, Ambiguous} can be triggered. For the latter two actions, large-scale web searches (Piktus et al., 2021; Komeili et al., 2022) are integrated as a strategic extension, since retrieval from static and limited corpora can only return sub-optimal documents in terms of scope and diversity. This augmentation is implemented to broaden the spectrum of retrieved information, harnessing the expansive and dynamic nature of the web to complement and enrich the initially obtained documents. Furthermore, to eliminate redundant contexts contained in retrieved documents that are unhelpful for RAG, a decompose-then-recompose algorithm is meticulously crafted throughout the retrieval and utilization process. This algorithm ensures the refinement of retrieved information, optimizing the extraction of key insights and minimizing the inclusion of non-essential elements, thereby enhancing the utilization of retrieved data.

CRAG is plug-and-play and experimentally implemented into RAG (Lewis et al., 2020) and Self-RAG (Asai et al., 2023) for demonstrating its adaptability to RAG-based approaches. Results on four datasets of PopQA (Mallen et al., 2023), Biography (Min et al., 2023), Pub Health (Zhang et al., 2023a), and Arc-Challenge (Bhakthavatsalam et al., 2021) show that CRAG can significantly improve the performance of standard RAG and state-of-the-art Self-RAG, demonstrating its generalizability across both short- and long-form generation tasks. To facilitate others to reproduce our results, we will publish all source code later.

In summary, our contributions in this paper are three-fold: 1) This paper studies the scenarios where the retriever returns inaccurate results and, to the best of our knowledge, makes the first attempt to design corrective strategies for RAG to improve its robustness. 2) A plug-and-play method named CRAG is proposed to improve the ability of automatic self-correction and efficient utilization of retrieved documents. 3) Experimental results extensively demonstrate CRAG’s adaptability to RAG-based approaches and its generalizability across short- and long-form generation tasks.

2 Related Work

Hallucinations of LLMs

Although LLMs have exhibited impressive abilities to understand instructions and generate fluent language texts (Bang et al., 2023; Qin et al., 2023; Zhong et al., 2023), one of the most severe issues that LLMs have still been struggling with is hallucinations. As many studies found (Zhang et al., 2023b; Shuster et al., 2021), either outdated information or incorrect knowledge that is activated would seriously result in hallucinations. Large-scale unregulated training data collection, low proportion of high-quality sampling data, imperfection of data allocation in the input space, and many other realistic factors could impact the LLMs and exacerbate the problems. Thus, it is obvious that the lack of accurate and specific knowledge can lead to misleading or even inaccurate generation, which will severely hurt the experience of users in most practical applications.

Retrieval-Augmented Generation

RAG (Lewis et al., 2020; Guu et al., 2020) is regarded as a useful method to address the issues above, which enhances the input questions of generative LMs with retrieved documents. It usually provides an extra knowledge source from a specific corpus, i.e., Wikipedia, which greatly improves the performance of LMs in a variety of tasks, especially in the knowledge-intensive ones. The proposed methods generally leverage information retrieval to supply documents containing relevant knowledge for generative LLMs. Earlier studies adopt either

Advanced RAG

Many advanced approaches have been developed from the original RAG in recent years. Considering that retrieval is sometimes unnecessary for some queries, conversely, responses without retrieval are even more accurate in many situations. Self-RAG (Asai et al., 2023) is proposed to selectively retrieve knowledge and introduce a critic model to decide whether to retrieve. Yoran et al. (2023) designed an NLI model to identify the irrelevant context and improve robustness. SAIL (Luo et al., 2023) is tuned on instructions to insert retrieved documents before instructions. While Toolformer (Schick et al., 2023) is pre-trained for calling APIs such as Wikipedia. In addition, in some long-text generation tasks, external knowledge is needed more than once, and when to retrieve should be concerned. Jiang et al. (2023) actively anticipate future content and decide when and what to retrieve in long-form generation.

Compared with recent studies (Schick et al., 2023; Luo et al., 2023; Asai et al., 2023) that are the most relevant to our work, a main difference should be highlighted. These approaches target on exploiting retrieval as a useful tool to augment generation or whether retrieval is necessary, while this study particularly studies the scenarios where the retriever returns inaccurate results. To the best of our knowledge, this paper makes the first attempt to explore and design corrective strategies for RAG to improve its robustness of generation.

3 Task Formulation

Following previous work (Lewis et al., 2020; Asai et al., 2023), given input X and an accessible corpus containing a large amount of knowledge documents C = {d1, ..., dN }, the system is expected to generate the output Y. The entire framework is usually divided into a retriever R and a generator G. The retriever R aims to retrieve the top-K documents D = {dr1 , ..., drk } that are relevant to the input X from the corpus C. Based on the input X and the retrieved results D, the generator G is responsible for generating the output Y. This framework can be formulated as:

P (Y|X ) = P (D|X )P (Y, D|X ). (1)

It shows that the retriever and generator are seamlessly coupled, exhibiting low risk tolerance. Any unsuccessful retrieval can result in an unsatisfactory response, regardless of the impressive abilities of the generator. This is exactly the focus of this paper to improve the robustness of generation.

4 CRAG

4.1 Overview of Model Inference

Figure 2 and Algorithm 1 present an overview of CRAG at inference, which designs corrective strategies to improve the robustness of generation. Given an input query and the retrieved documents from any retriever, a lightweight retrieval evaluator is constructed to estimate the relevance score of retrieved documents to the input query (Section 4.2). The relevance score is quantified into a total of three confidence degrees and then triggered the corresponding actions: {Correct, Incorrect, Ambiguous} (Section 4.3). If the action Correct is triggered, the retrieved documents will be refined into more precise knowledge strips. This refinement operation involves knowledge decomposition, filter, and recomposition (Section 4.4). If the action Incorrect is triggered, the retrieved documents will be discarded. Instead, web searches are resorted to and regarded as complementary knowledge sources for corrections (Section 4.5). Eventually, when it cannot confidently make a correct or incorrect judgment, a soft and balanced action Ambiguous which combines both of them is triggered. After optimizing the retrieval results, an arbitrary generative model can be adopted.

4.2 Retrieval Evaluator

It is natural to wonder whether the retrieved documents are accurate or not before using them, which is significant since irrelevant or misleading messages can be identified in this way. The accuracy of the retrieval evaluator undeniably plays a pivotal role in shaping the overall system performance, as it influences the outcomes of subsequent processes. Our objective is to correct the retrieved documents if they are irrelevant. Specifically, T5-large (Raffel et al., 2020) is adopted for initializing the retrieval evaluator and fine-tuned. The relevance signals for fine-tuning the evaluator can be collected from the existing datasets. More details about this fine-tuning step can be referred to in Appendix B.2.

Retrieval

x: Who was the screenwriter for Death of a Batman?

Retrieved Documents

d1 d2

Retrieval

Ask: If retrieved documents are correct to x?

d1	strip1	strip1
	strip2		kin
	…	stripk
d2	Decompose	Filter	Recompose
		stripk

Knowledge Correction

Correct

Knowledge Searching

Ambiguous	x	q: Death of a Batman;	k1	k2	kex
Incorrect	Rewrite	Web	Select	Search	kn

Generation

Figure 2: An overview of CRAG at inference. A retrieval evaluator is constructed to evaluate the relevance of the retrieved documents to the input, and estimate a confidence degree based on which different knowledge retrieval actions of {Correct, Incorrect, Ambiguous} can be triggered.

Every question, there are generally 10 documents retrieved. The question is concatenated with each single document as the input, and the evaluator predicts the relevance score for each question-document pair individually. We also tried to prompt ChatGPT to identify the retrieval relevance for comparison, but it underperforms as elaborated in Section 5.5. Based on these calculated relevance scores, a final judgment is made as to whether the retrieval is correct or not associated with the action trigger. Compared with the critic model of Self-RAG (Asai et al., 2023) that instruction-tuned LLaMA-2 (7B), the evaluator designed in CRAG demonstrates the advantages of being quite lightweight (0.77B).

4.3 Action Trigger

To correct the irrelevant documents and refine the target documents as needed, actions should be executed discriminately. Based on the aforementioned confidence score for each retrieved document, three types of actions are designed and triggered accordingly where the upper and lower thresholds are set. If the confidence score is higher than the upper threshold, the retrieved document is identified as Correct, while identified as Incorrect if below the lower threshold. Otherwise, Ambiguous is executed. Each retrieved document is conducted individually and integrated eventually.

Correct

Here, a retrieval is assumed Correct when the confidence score of at least one retrieved document is higher than the upper threshold. If so, it means that there are relevant documents in the retrieved results. Even if a relevant document can be found, there is inevitably some noisy knowledge strips in this document. To extract the most critical knowledge strips within this document, a knowledge refinement method is further designed which will be elaborated in Section 4.4.

Incorrect

Besides, a retrieval is assumed Incorrect when the confidence scores of all retrieved documents are below the lower threshold. This indicates that all retrieved documents are considered irrelevant, which are unhelpful for generation. Therefore, we need to seek new sources of knowledge for correction. Here, web search is introduced to search from the Internet as

Algorithm 1: CRAG Inference

Require: E (Retrieval Evaluator), W (Query Rewriter), G (Generator)

Input: x (Input question), D = {d1, d2, ..., dk} (Retrieved documents)

Output: y (Generated response)

scorei = E evaluates the relevance of each pair (x, di), di ∈ D
Confidence = Calculate and give a final judgment based on {score1, score2, ...scorek}
// Confidence has 3 optional values: [CORRECT], [INCORRECT] or [AMBIGUOUS]
if Confidence == [CORRECT] then
Internal_Knowledge = Knowledge_Refine(x, D)
k = Internal_Knowledge
else if Confidence == [INCORRECT] then
External_Knowledge = Web_Search(W Rewrites x for searching)
k = External_Knowledge
else if Confidence == [AMBIGUOUS] then
Internal_Knowledge = Knowledge_Refine(x, D)
External_Knowledge = Web_Search(W Rewrites x for searching)
k = Internal_Knowledge + External_Knowledge
end
G predicts y given x and k

elaborated in Section 4.5. This corrective action helps overcome the embarrassing challenge where no reliable knowledge can be referred to.

Ambiguous

Except for the above two situations, the remaining will be assigned to an intermediate action of Ambiguous. Since the retrieval evaluator is not confident in its judgment, both types of processed knowledge in Correct and Incorrect are combined to complement each other. Implementing such a moderating and soft strategy can significantly contribute to strengthening the robustness and resilience of the system, fostering a more adaptable framework for optimal performance.

4.4 Knowledge Refinement

Given a retrieved relevant document, a decompose-then-recompose knowledge refinement method is designed to further extract the most critical knowledge strips in it. First, each retrieved document is segmented into fine-grained knowledge strips through heuristic rules, more details are available in Appendix B.2. Then, the retrieval evaluator fine-tuned in Section 4.2 is employed to calculate the relevance score of each knowledge strip. Based on these scores, irrelevant knowledge strips are filtered out, while relevant ones are recomposed via concatenation in order, namely internal knowledge.

4.5 Web Search

It is extremely important to seek complementary external knowledge if the retrieved results are all assumed irrelevant. Since retrieval from static and limited corpora can only return sub-optimal documents in terms of scope and diversity, large-scale web searches (Piktus et al., 2021; Komeili et al., 2022) are integrated as a strategic extension of RAG. Specifically, the inputs are rewritten into queries composed of keywords by ChatGPT to mimic the daily usage of search engine. The prompt for rewriting is shown in Appendix A. In CRAG, a public and accessible commercial web search API is adopted to generate a series of URL links for every query.2 Moreover, we utilize the URL links to navigate web pages, transcribe their content, and employ the same knowledge refinement method as Section 4.4 to derive the relevant web knowledge, namely external knowledge.

5 Experiments

We conducted experiments to extensively demonstrate CRAG’s adaptability to RAG-based approaches and its generalizability across both short- and long-form generation tasks.

5.1 Tasks, Datasets and Metrics

CRAG was evaluated on four datasets, including PopQA (Mallen et al., 2023) (short-form generation), Biography (Min et al., 2023) (long-form generation), PubHealth (Zhang et al., 2023a) (true-or-false question), and Arc-Challenge (Bhakthavatsalam et al., 2021) (multiple-choice question).

2In this study, Google Search API is utilized for searching.

Table 1: Overall evaluation results on the test sets of four datasets.

Method	PopQA (Accuracy)	Bio (FactScore)	Pub (Accuracy)	ARC (Accuracy)
LMs trained with propriety data
LLaMA2-c13B	20.0	55.9	49.4	38.4
Ret-LLaMA2-c13B	51.8	79.9	52.1	37.9
ChatGPT	29.3	71.8	70.1	75.3
Ret-ChatGPT	50.8	-	54.7	75.3
Perplexity.ai	-	71.2	-	-
Baselines without retrieval
LLaMA27B	14.7	44.5	34.2	21.8
Alpaca7B	23.6	45.8	49.8	45.0
LLaMA213B	14.7	53.4	29.4	29.4
Alpaca13B	24.4	50.2	55.5	54.9
CoVE65B	-	71.2	-	-
Baselines with retrieval
LLaMA27B	38.2	78.0	30.0	48.0
Alpaca7B	46.7	76.6	40.2	48.0
SAIL	-	-	69.2	48.4
LLaMA213B	45.7	77.5	30.2	26.0
Alpaca13B	46.1	77.7	51.1	57.6
LLaMA2-hf-7b
RAG	37.7	44.9	9.1	23.8
CRAG	39.8	47.7	9.1	25.8
Self-RAG*	29.0	32.2	0.7	23.9
Self-CRAG	49.0	69.1	0.6	27.9
SelfRAG-LLaMA2-7b
RAG	40.3	59.2	39.0	46.7
CRAG	59.3	74.1	75.6	54.8
Self-RAG	54.9	81.2	72.4	67.3
Self-CRAG	61.8	86.2	74.8	67.2

Bold numbers indicate the best performance among all methods and LLMs. Gray-colored bold scores indicate the best performance using a specific LLM. * indicates the results reproduced by us, otherwise results except ours are cited from their original papers.

Following previous work, accuracy was adopted as the evaluation metric for PopQA, PubHealth, and Arc-Challenge. FactScore (Min et al., 2023) was adopted as the evaluation metric for Biography. Readers can refer to Appendix B.1 for more details.

5.2 Baselines

We primarily compared CRAG with both approaches without and with retrieval, where the latter consists of standard RAG and advanced RAG.

Baselines without retrieval

We evaluated some public LLMs, LLaMA2-7B,13B (Touvron et al., 2023b), instruction-tuned models, Alpaca-7B,13B (Dubois et al., 2023), and CoVE65B (Dhuliawala et al., 2023) which introduces iterative engineering to improve the factuality of LLM generations. Propriety LLMs such as LLaMA2-chat13B and ChatGPT are also included.

Advanced RAG

(1) SAIL (Luo et al., 2023) that instruction-tuned an LM on the Alpaca instruction-tuning data with top retrieved documents inserted before instructions. (2) Self-RAG (Asai et al., 2023) that tuned the LLaMA2 on the instruction-tuning data containing several sets of reflection tokens which were labeled by GPT-4 (OpenAI, 2023). (3) Following Asai et al. (2023), we also.

5.3 Results

Table 1 presents the results on four datasets. The model coupling the proposed method with standard RAG is named CRAG and that coupling with Self-RAG is named Self-CRAG. Readers can refer to Appendix B.2 for more implementation details of our proposed methods. From these results, we can conclude the following findings:

First, the proposed method can significantly improve the performance of RAG and Self-RAG. Specifically, CRAG outperformed RAG by margins of 19.0% accuracy on PopQA, 14.9% FactScore on Biography, 36.6% accuracy on PubHealth, and 8.1% accuracy on Arc-Challenge when based on SelfRAG-LLaMA2-7b, as well as by margins of 2.1% accuracy on PopQA, 2.8% FactScore on Biography, and 2.0% on Arc-Challenge when based on LLaMA2-hf-7b. Compared with the current state-of-the-art Self-RAG, Self-CRAG outperformed it by margins of 20.0% accuracy on PopQA, 36.9% FactScore on Biography, and 4.0% accuracy on Arc-Challenge when based on LLaMA2-hf-7b, as well as by margins of 6.9% accuracy on PopQA, 5.0% FactScore on Biography, and 2.4% accuracy on PubHealth, when based on SelfRAG-LLaMA2-7b. These results demonstrated the adaptability of CRAG which is plug-and-play and can be implemented into RAG-based approaches.

Second, the proposed method demonstrated great generalizability across a variety of generation tasks. In particular, these benchmarks reported in Table 1 respectively represent different practical scenarios including short-form entity generation (PopQA), long-form generation (Biography), and closed-set tasks (PubHealth, Arc-Challenge). These results verified the consistent effectiveness of CRAG. Its versatility across a spectrum of tasks underscores its robust capabilities and generalizability across diverse scenarios.

Third, the proposed method exhibited greater flexibility in replacing the underlying LLM generator. It can be seen that CRAG still showed competitive performance when the underlying LLMs was changed from SelfRAG-LLaMA2-7b to LLaMA2-hf-7b, while the performance of Self-RAG dropped significantly, even underperforming the standard RAG on several benchmarks. The reason for these results is that Self-RAG needs to be instruction-tuned using human or LLM annotated data to learn to output special critic tokens as needed, while this ability is not learned in common LLMs. CRAG does not have any requirements for this ability. As you can imagine, when more advanced LLMs are available in the future, they can be coupled with CRAG easily, while additional instruction tuning is still necessary for Self-RAG.

5.4 Ablation Study

The impact of each triggered action. To further verify the effectiveness of triggered actions designed in the retrieval evaluator, ablation tests for removing each single action in the proposed method were conducted as shown in Table 2. Evaluations on the PopQA dataset were conducted to demonstrate the performance change in terms of accuracy. Specifically, when the action Correct or Incorrect was removed, it was merged with Ambiguous so that the proportion that originally triggered Correct or Incorrect would trigger Ambiguous. On the other hand, when the action Ambiguous was removed, there was only one threshold against which all input queries clearly triggered Correct or Incorrect. From these results, it can be seen that there was a performance drop no matter which action was removed, illustrating that each action contributed to improving the robustness of generation.

The impact of each knowledge utilization operation. Table 3 illustrated how the performance changed if a key knowledge utilization operation was ablated. Evaluations on the PopQA dataset in terms of accuracy were conducted by individually removing the knowledge utilization operations of document refinement, search query rewriting, and

Model	Self-LLaMA2-hf-7b	SelfRAG-LLaMA2-7b
CRAG	w/o. Correct	w/o. Incorrect	w/o. Ambiguous
CRAG	47.3	44.5	46.8	45.7
Self-CRAG	49.0	43.6	47.7	48.1

Table 2: Ablation study for removing each single action on the PopQA dataset in terms of accuracy.

5.5 Accuracy of the Retrieval Evaluator

The quality of the retrieval evaluator significantly determined the performance of the entire system. Given the document retrieval results, we assessed whether the retrieval evaluator can accurately determine the overall quality of these results. The assessment accuracy on the PopQA dataset of our retrieval evaluator and the commercial LLM ChatGPT on the document retrieval results was shown in Table 4. The prompts of ChatGPT, ChatGPT-CoT, and ChatGPT-few-shot used in our experiments can be referred to in Appendix A. Results reveal that the lightweight T5-based retrieval evaluator significantly outperformed the competitive ChatGPT in all settings.

5.6 Robustness to Retrieval Performance

To further verify the robustness of the proposed method to retrieval performance, we studied how the generation performance changed given different retrieval performance. A part of accurate retrieval results were deliberately removed at random to imitate a low-quality retriever and evaluate how the performance changed. Figure 3 demonstrated the performance change of Self-RAG and Self-CRAG on the PopQA dataset. It can be seen that the generation performance of Self-RAG and Self-CRAG dropped as the retrieval performance dropped, indicating that the generator relied heavily on the quality of the retriever. Furthermore, as the retrieval performance dropped, the generation performance of Self-CRAG dropped more slightly than that of Self-RAG. These results imply the superiority of Self-CRAG over Self-RAG on enhancing the robustness to retrieval performance.

6 Conclusion

This paper studies the problem where RAG-based approaches are challenged if retrieval goes wrong, thereby exposing inaccurate and misleading knowledge to generative LMs. Corrective Retrieval Augmented Generation is proposed to improve the robustness of generation. Essentially, a lightweight retrieval evaluator is to estimate and trigger three knowledge retrieval actions discriminately. With the further leverage of web search and optimized knowledge utilization, CRAG has significantly improved the ability of automatic self-correction and efficient utilization of retrieved documents. Experiments extensively demonstrate its adaptability to RAG-based approaches as well as generalizability across short- and long-form generation tasks.

Table 3: Ablation study for removing each knowledge utilization operation on the PopQA in terms of accuracy.

	LLaMA2-hf-7b	SelfRAG-LLaMA2-7b
CRAG	47.3	59.3
w/o. refinement	38.9	47.0
w/o. rewriting	44.8	56.6
w/o. selection	44.0	53.8
Self-CRAG	49.0	61.8
w/o. refinement	35.9	52.2
w/o. rewriting	37.2	58.4
w/o. selection	24.9	57.9

Table 4: Evaluation of our retrieval evaluator and ChatGPT for the retrieval results on the PopQA dataset.

	Accuracy
Our Retrieval Evaluator (T5-based)	84.3
ChatGPT	58.0
ChatGPT-CoT	62.4
ChatGPT-few-shot	64.7

Figure 3: The generation performance of Self-RAG, Self-CRAG, RAG, and CRAG given different retrieval performance on the PopQA dataset with SelfRAG-LLaMA-7b.

The lower horizontal line demonstrates the performance of the generator without retrieval.

Limitations

While we primarily proposed to improve the RAG framework from a corrective perspective, how to detect and correct the wrong knowledge more accurately and effectively still requires further study. Although CRAG can be seamlessly coupled with various RAG-based approaches, fine-tuning a retrieval evaluator is inevitable. In addition, potential bias introduced by web searches is also worth concern. The quality of internet sources can vary significantly, and incorporating such data without enough consideration may introduce noise or misleading information to the generated outputs.

Future work will further explore a more stable and reliable method of retrieval augmentation.

References

Rohan Anil, Andrew M. Dai, Orhan Firat, Melvin Johnson, Dmitry Lepikhin, Alexandre Passos, Siamak Shakeri, Emanuel Taropa, Paige Bailey, Zhifeng Chen, Eric Chu, Jonathan H. Clark, Laurent El Shafey, Yanping Huang, Kathy Meier-Hellstern, Gaurav Mishra, Erica Moreira, Mark Omernick, Kevin Robinson, Sebastian Ruder, et al. 2023. PaLM 2 technical report. CoRR, abs/2305.10403.

Akari Asai, Zeqiu Wu, Yizhong Wang, Avirup Sil, and Hannaneh Hajishirzi. 2023. Self-rag: Learning to retrieve, generate, and critique through self-reflection. CoRR, abs/2310.11511.

Yejin Bang, Samuel Cahyawijaya, Nayeon Lee, Wenliang Dai, Dan Su, Bryan Wilie, Holy Lovenia, Ziwei Ji, Tiezheng Yu, Willy Chung, Quyet V. Do, Yan Xu, and Pascale Fung. 2023. A multitask, multilingual, multimodal evaluation of chatgpt on reasoning, hallucination, and interactivity. CoRR, abs/2302.04023.

Sumithra Bhakthavatsalam, Daniel Khashabi, Tushar Khot, Bhavana Dalvi Mishra, Kyle Richardson, Ashish Sabharwal, Carissa Schoenick, Oyvind Tafjord, and Peter Clark. 2021. Think you have solved direct-answer question answering? try arc-da, the direct-answer AI2 reasoning challenge. CoRR, abs/2102.03315.

Tom B Brown, Benjamin Mann, Nick Ryder, et al. 2020. Language models are few-shot learners. In Advances in neural information processing systems, pages 1877–1901.

Aakanksha Chowdhery, Sharan Narang, Jacob Devlin, Maarten Bosma, Gaurav Mishra, Adam Roberts, Paul Barham, Hyung Won Chung, Charles Sutton, Sebastian Gehrmann, Parker Schuh, Kensen Shi, Sasha Tsvyashchenko, Joshua Maynez, Abhishek Rao, Parker Barnes, Yi Tay, Noam Shazeer, Vinodkumar Prabhakaran, Emily Reif, Nan Du, Ben Hutchinson, Reiner Pope, James Bradbury, Jacob Austin, Michael Isard, Guy Gur-Ari, Pengcheng Yin, Toju Duke, Anselm Levskaya, Sanjay Ghemawat, Sunipa Dev, Henryk Michalewski, Xavier Garcia, Vedant Misra, Kevin Robinson, Liam Fedus, Denny Zhou, Daphne Ippolito, David Luan, Hyeontaek Lim, Barret Zoph, Alexander Spiridonov, Ryan Sepassi, David Dohan, Shivani Agrawal, Mark Omernick, Andrew M. Dai, Thanumalayan Sankaranarayana Pillai, Marie Pellat, Aitor Lewkowycz, Erica Moreira, Rewon Child, Oleksandr Polozov, Katherine Lee, Zongwei Zhou, Xuezhi Wang, Brennan Saeta, Mark Diaz, Orhan Firat, Michele Catasta, Jason Wei, Kathy Meier-Hellstern, Douglas Eck, Jeff Dean, Slav Petrov, and Noah Fiedel. 2023. Palm: Scaling language modeling with pathways. J. Mach. Learn. Res., 24:240:1–240:113.

Shehzaad Dhuliawala, Mojtaba Komeili, Jing Xu, Roberta Raileanu, Xian Li, Asli Celikyilmaz, and Jason Weston. 2023. Chain-of-verification reduces hallucination in large language models. CoRR, abs/2309.11495.

Yann Dubois, Xuechen Li, Rohan Taori, Tianyi Zhang, Ishaan Gulrajani, Jimmy Ba, Carlos Guestrin, Percy Liang, and Tatsunori B. Hashimoto. 2023. Alpaca-farm: A simulation framework for methods that learn from human feedback. CoRR, abs/2305.14387.

Kelvin Guu, Kenton Lee, Zora Tung, Panupong Pasupat, and Ming-Wei Chang. 2020. Retrieval augmented language model pre-training. In Proceedings of the 37th International Conference on Machine Learning, ICML 2020, 13-18 July 2020, Virtual Event, volume 119 of Proceedings of Machine Learning Research, pages 3929–3938. PMLR.

Gautier Izacard, Mathilde Caron, Lucas Hosseini, Sebastian Riedel, Piotr Bojanowski, Armand Joulin, and Edouard Grave. 2022. Unsupervised dense information retrieval with contrastive learning. Trans. Mach. Learn. Res., 2022.

Ziwei Ji, Nayeon Lee, Rita Frieske, Tiezheng Yu, Dan Su, Yan Xu, Etsuko Ishii, Yejin Bang, Andrea Madotto, and Pascale Fung. 2023. Survey of hallucination in natural language generation. ACM Comput. Surv., 55(12):248:1–248:38.

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. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, EMNLP 2023, Singapore, December 6-10, 2023, pages 7969–7992. Association for Computational Linguistics.

Mojtaba Komeili, Kurt Shuster, and Jason Weston. 2022. Internet-augmented dialogue generation. In Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), ACL 2022, Dublin, Ireland, May 22-27, 2022, pages 8460–8478. Association for Computational Linguistics.

References

Patrick S. H. Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich Küttler, Mike Lewis, Wen-tau Yih, Tim Rocktäschel, Sebastian Riedel, and Douwe Kiela. 2020. Retrieval-augmented generation for knowledge-intensive NLP tasks. In Advances in Neural Information Processing Systems 33: Annual Conference on Neural Information Processing Systems 2020, NeurIPS 2020, December 6-12, 2020, virtual.

Huayang Li, Yixuan Su, Deng Cai, Yan Wang, and Lemao Liu. 2022. A survey on retrieval-augmented text generation. CoRR, abs/2202.01110.

Hongyin Luo, Tianhua Zhang, Yung-Sung Chuang, Yuan Gong, Yoon Kim, Xixin Wu, Helen Meng, and James R. Glass. 2023. Search augmented instruction learning. In Findings of the Association for Computational Linguistics: EMNLP 2023, Singapore, December 6-10, 2023, pages 3717–3729. Association for Computational Linguistics.

Alex Mallen, Akari Asai, Victor Zhong, Rajarshi Das, Daniel Khashabi, and Hannaneh Hajishirzi. 2023. When not to trust language models: Investigating effectiveness of parametric and non-parametric memories. In Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), ACL 2023, Toronto, Canada, July 9-14, 2023, pages 9802–9822. Association for Computational Linguistics.

Sewon Min, Kalpesh Krishna, Xinxi Lyu, Mike Lewis, Wen-tau Yih, Pang Wei Koh, Mohit Iyyer, Luke Zettlemoyer, and Hannaneh Hajishirzi. 2023. Factscore: Fine-grained atomic evaluation of factual precision in long form text generation. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, EMNLP 2023, Singapore, December 6-10, 2023, pages 12076–12100. Association for Computational Linguistics.

Dor Muhlgay, Ori Ram, Inbal Magar, Yoav Levine, Nir Ratner, Yonatan Belinkov, Omri Abend, Kevin Leyton-Brown, Amnon Shashua, and Yoav Shoham. 2023. Generating benchmarks for factuality evaluation of language models. CoRR, abs/2307.06908.

OpenAI. 2023. GPT-4 technical report. CoRR, abs/2303.08774.

Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Carroll L. Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, John Schulman, Jacob Hilton, Fraser Kelton, Luke Miller, Maddie Simens, Amanda Askell, Peter Welinder, Paul F. Christiano, Jan Leike, and Ryan Lowe. 2022. Training language models to follow instructions with human feedback. In NeurIPS.

Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Dmytro Okhonko, Samuel Broscheit, Gautier Izacard, Patrick S. H. Lewis, Barlas Oguz, Edouard Grave, Wen-tau Yih, and Sebastian Riedel. 2021. The web is your oyster - knowledge-intensive NLP against a very large web corpus. CoRR, abs/2112.09924.

Chengwei Qin, Aston Zhang, Zhuosheng Zhang, Jiaao Chen, Michihiro Yasunaga, and Diyi Yang. 2023. Is chatgpt a general-purpose natural language processing task solver? In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, EMNLP 2023, Singapore, December 6-10, 2023, pages 1339–1384. Association for Computational Linguistics.

Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J. Liu. 2020. Exploring the limits of transfer learning with a unified text-to-text transformer. J. Mach. Learn. Res., 21:140:1–140:67.

Md. Rashad Al Hasan Rony, Ricardo Usbeck, and Jens Lehmann. 2022. Dialokg: Knowledge-structure aware task-oriented dialogue generation. In Findings of the Association for Computational Linguistics: NAACL 2022, Seattle, WA, United States, July 10-15, 2022, pages 2557–2571. Association for Computational Linguistics.

Timo Schick, Jane Dwivedi-Yu, Roberto Dessì, Roberta Raileanu, Maria Lomeli, Luke Zettlemoyer, Nicola Cancedda, and Thomas Scialom. 2023. Toolformer: Language models can teach themselves to use tools. CoRR, abs/2302.04761.

Freda Shi, Xinyun Chen, Kanishka Misra, Nathan Scales, David Dohan, Ed H. Chi, Nathanael Schärli, and Denny Zhou. 2023. Large language models can be easily distracted by irrelevant context. In Proceedings of the 40th International Conference on Machine Learning, volume 202 of Proceedings of Machine Learning Research, pages 31210–31227. PMLR.

Kurt Shuster, Spencer Poff, Moya Chen, Douwe Kiela, and Jason Weston. 2021. Retrieval augmentation reduces hallucination in conversation. In Findings of the Association for Computational Linguistics: EMNLP 2021, Virtual Event / Punta Cana, Dominican Republic, 16-20 November, 2021, pages 3784–3803. Association for Computational Linguistics.

Chao-Hong Tan, Jia-Chen Gu, Chongyang Tao, Zhen-Hua Ling, Can Xu, Huang Hu, Xiubo Geng, and Daxin Jiang. 2022. Tegtok: Augmenting text generation via task-specific and open-world knowledge. In Findings of the Association for Computational Linguistics: ACL 2022, Dublin, Ireland, May 22-27, 2022, pages 1597–1609. Association for Computational Linguistics.

Hugo Touvron, Thibaut Lavril, Gautier Izacard, Xavier Martinet, Marie-Anne Lachaux, Timothée Lacroix, Baptiste Rozière, Naman Goyal, Eric Hambro, Faisal Azhar, Aurélien Rodriguez, Armand Joulin, Edouard Grave, and Guillaume Lample. 2023a. Llama: Open, and efficient foundation language models. CoRR abs/2302.13971.

References

Hugo Touvron, Louis Martin, Kevin Stone, Peter Albert, Amjad Almahairi, Yasmine Babaei, Nikolay Bashlykov, Soumya Batra, Prajjwal Bhargava, Shruti Bhosale, Dan Bikel, Lukas Blecher, et al. 2023b. Llama 2: Open foundation and fine-tuned chat models. CoRR, abs/2307.09288.

Ori Yoran, Tomer Wolfson, Ori Ram, and Jonathan Berant. 2023. Making retrieval-augmented language models robust to irrelevant context. CoRR, abs/2310.01558.

Tianhua Zhang, Hongyin Luo, Yung-Sung Chuang, Wei Fang, Luc Gaitskell, Thomas Hartvigsen, Xixin Wu, Danny Fox, Helen Meng, and James R. Glass. 2023a. Interpretable unified language checking. CoRR, abs/2304.03728.

Yue Zhang, Yafu Li, Leyang Cui, Deng Cai, Lemao Liu, Tingchen Fu, Xinting Huang, Enbo Zhao, Yu Zhang, Yulong Chen, Longyue Wang, Anh Tuan Luu, Wei Bi, Freda Shi, and Shuming Shi. 2023b. Siren’s song in the AI ocean: A survey on hallucination in large language models. CoRR, abs/2309.01219.

Qihuang Zhong, Liang Ding, Juhua Liu, Bo Du, and Dacheng Tao. 2023. Can chatgpt understand too? A comparative study on chatgpt and fine-tuned BERT. CoRR, abs/2302.10198.

Task Prompts

The prompts for generating knowledge keywords as web search queries were illustrated in Table 5.

Extract at most three keywords separated by comma from the following dialogues and questions as queries for the web search, including topic background within dialogues and main intent within questions.

question: What is Henry Feilden’s occupation?	query: Henry Feilden, occupation
question: In what city was Billy Carlson born?	query: city, Billy Carlson, born
question: What is the religion of John Gwynn?	query: religion of John Gwynn
question: What sport does Kiribati men’s national basketball team play?	query: sport, Kiribati men’s national basketball team play
question: [question]	query:

Table 5: The few-shot prompt to GPT-3.5 Turbo for generating knowledge keywords as web search queries.

The prompts to instruct ChatGPT as the evaluator were illustrated in Table 6, Table 7, and Table 8 respectively.

Given a question, does the following document have exact information to answer the question? Answer yes or no only.

Question: [question]

Document: [document]

Given a question, does the following document have exact information to answer the question? Answer yes or no only.

Question: In what city was Abraham Raimbach born?

Document: Bancroft was born on November 25, 1839 in New Ipswich, New Hampshire to James Bancroft and Sarah Kimball. At an early age he was cared for by Mr. and Mrs. Patch of Ashby, Massachusetts, the neighboring town. While not legally adopted, they named him Cecil Franklin Patch Bancroft, adding Franklin Patch after the son Mr. and Mrs. Patch had who recently died. He attended public schools in Ashby as well as the Appleton Academy in New Ipswich. He entered Dartmouth College in 1856 at the age of sixteen and graduated in 1860 near the top of his class. Bancroft continued his education as he began his career in teaching. He took classes at the Union Theological Seminary in New York City during the 1864-65 academic year. While there he was a member of the United States Christian Commission, traveling to support soldiers during the Civil War. He then transferred to the Andover Theological Seminary where he would graduate in 1867.

Answer: No.

Question: In what country is Wilcza Jama, Sokółka County?

Document: Wilcza Jama is a village in the administrative district of Gmina Sokółka, within Sokółka County, Podlaskie Voivodeship, in north-eastern Poland, close to the border with Belarus.

Answer: Yes.

Question: What sport does 2004 Legg Mason Tennis Classic play?

Document: The 2004 Legg Mason Tennis Classic was the 36th edition of this tennis tournament and was played on outdoor hard courts. The tournament was part of the International Series of the 2004 ATP Tour. It was held at the William H.G. FitzGerald Tennis Center in Washington, D.C. from August 16 through August 22, 2004.

Answer: Yes.

Table 6: The direct prompt to GPT-3.5 Turbo as the evaluator.

Given a question, does the following document have exact information to answer the question?

Question: [question]

Document: [document]

Think Step by step, and answer with yes or no only.

Table 7: The prompt to GPT-3.5 Turbo with Chain-of-Thought as the evaluator.

Question: Who is the author of Skin?

Document: The Skin We’re In: A Year of Black Resistance and Power is a book by Desmond Cole published by Doubleday Canada in 2020. The Skin We’re In describes the struggle against racism in Canada during the year 2017, chronicling Cole’s role as an anti-racist activist and the impact of systemic racism in Canadian society. Among the events it discusses are the aftermath of the assault of Dafonte Miller in late 2016 and Canada 150. The work argues that Canada is not immune to the anti-Black racism that characterizes American society. Due to an error by the publisher, the initial printing of the book’s cover did not include word Black in the subtitle. The mistake was later corrected. The book won the Toronto Book Award for 2020. In 2021, the book was nominated for the Shaughnessy Cohen Prize for Political Writing.

Answer: No.

Question: [question]

Document: [document]

Answer:

Table 8: The few-shot prompt to GPT-3.5 Turbo as the evaluator.

Experiments

Tasks, Datasets and Metrics

CRAG was evaluated on four datasets, which are in public domain and licensed for research purposes, including:

PopQA (Mallen et al., 2023) is a short-form generation task. Generally, only one entity of factual knowledge is expected to be answered for each single question. In our experiments, we exactly followed the setting in Self-RAG (Asai et al., 2023) which evaluated methods on a long-tail subset consisting of 1,399 rare entity queries whose monthly Wikipedia page views are less than 100. Accuracy was adopted as the evaluation metric.
Biography (Min et al., 2023) is a long-form generation task that is tasked to generate a detailed biography about a certain entity. Following previous work, FactScore (Min et al., 2023) was adopted to evaluate the generated biographies.
PubHealth (Zhang et al., 2023a) is a task in health care domain consisting of true-or-false questions. Claims are represented about health with factual information, and the model is tasked to verify the authenticity and give the judgment. Accuracy was adopted as the evaluation metric.
Arc-Challenge (Bhakthavatsalam et al., 2021) is a multiple-choice question task about some daily commonsense science phenomena. Given a scientific event that occurs in daily life, the model is required to select the correct description among 3 or 4 optional choices. Accuracy was adopted as the evaluation metric as well.

Implementation Details

Retrieval Evaluator: We fine-tuned the retrieval evaluator based on the lightweight T5-large (Raffel et al., 2020) pre-trained model. Its parameter size is much smaller than the most current LLMs (Touvron et al., 2023a,b; Chowdhery et al., 2023; Anil et al., 2023; Brown et al., 2020; Ouyang et al., 2022; OpenAI, 2023). To ensure all experimental results were comparable with Self-RAG (Asai et al., 2023), the same retrieval results through Contriver (Izacard et al., 2022) were provided by Self-RAG and were also adopted in our experiments. The relevance signals for fine-tuning the evaluator can be collected from the existing datasets. For example, PopQA (Mallen et al., 2023) provides the golden subject wiki title from Wikipedia for each question. We can use that to track a not 100% relevant but rather high-quality passage. We utilized that as the relevance labels for fine-tuning the retrieval evaluator.3 On the other hand, the negative samples were randomly sampled and we used the version provided by Self-RAG (Asai et al., 2023). Specifically, the original PopQA dataset consists of 14k samples, 1,399 of which were used for testing following Self-RAG (Asai et al., 2023), and the remaining were used for fine-tuning to avoid information leakage. Besides, the fine-tuned evaluator was transferred and also utilized on the Bio, Pub and ARC datasets during inference. The label of positive samples was 1, while that of negative ones was -1. At inference, the evaluator scored the relevance from -1 to 1 for each document. The two confidence thresholds for triggering one of the three actions were set empirically. Specifically, they were set as (0.59, -0.99) in PopQA, (0.5, -0.91) in PubQA and Arc-Challenge, as well as (0.95, -0.91) in Biography.

Internal Knowledge: To obtain fine-grained retrieval results, we segmented the retrieved results into internal strips. If a retrieved result is as short as one or two sentences, it is regarded as an individual strip, otherwise, retrieval documents are required to be split into smaller units which generally consist of a few sentences according to the total length. The scale is assumed to include an independent piece of information, and the filtering is based on the segments. We directly adopted the evaluator again for knowledge strips filtering, and the top-k is set to 5, filter threshold as -0.5.

External Knowledge: Google Search API was adopted to search for the relevant URLs, top-k is set to 5, and pages from Wikipedia will be added preferentially. The searched web pages are generally in the form of HTML files, where content is split with special tokens like <p> and </p>. Thus an extra segmentation like the knowledge refinement is not required, related knowledge paragraphs can be directly selected with the evaluator similar to internal knowledge.

Generator: As CRAG is a plug-and-play method, all generation models that can be utilized in RAG fit our approach as well. To be consistent with baselines for comparison, we adopted LLaMA2 (Touvron et al., 2023b) for the generation. We first introduced the LLaMA2-hf-7b from huggingface to generate responses. Since Self-RAG (Asai et al., 2023) fine-tuned LLaMA2 and reached a new state-of-the-art performance.

3https://huggingface.co/datasets/akariasai/PopQA

B.3 Results on PubHealth and Arc-Challenge

It is worth mentioning that the performance on PubHealth based on LLaMA2-hf-7b was much worse than others. We studied these cases and found that LLaMA2-hf-7b is relatively weak in instruction comprehension. Most of the cases fail to generate True or False in such a binary-question task, resulting in a low accuracy during the evaluation. This situation somewhat happens in Arc-Challenge as well, when the model is tasked to generate the index of a candidate.

Self-CRAG

To demonstrate that our plug-and-play approach can be utilized in other concurrent studies, we specifically designed to insert our CRAG into the Self-RAG (Asai et al., 2023) framework and named it Self-CRAG. Self-RAG is an advanced RAG approach that introduces a critic model to decide whether to retrieve and which retrieved document to be referred for generation. It meets our demand for deciding which action to be triggered, thus we replaced the retrieved items in Self-RAG with our processed internal knowledge for Correct, external knowledge for Incorrect, and combined knowledge for Ambiguous.

On several tasks, we further utilized the launched model, SelfRAG-LLaMA2-7b, as a new generator to be consistent with their work and study the specific improvement of our method.