IT-RAG-Bench: Italian Retrieval & RAG Benchmark

IT-RAG-Bench is a synthetic Italian-language retrieval benchmark designed to evaluate dense embedding models on document retrieval and Retrieval-Augmented Generation (RAG) tasks in Italian.

This dataset is the companion resource for the paper:

Benchmarking Google Embeddings 2 against Open-Source Models for Multilingual Dense Retrieval and RAG Systems
Stefano Cirillo, Domenico Desiato, Giuseppe Polese, Giandomenico Solimando — arXiv 2026
📄 https://arxiv.org/abs/2605.23618
💻 https://github.com/cciro94/GoogleEmbeddings2-benchmark

---

Dataset Summary

IT-RAG-Bench provides 3,200 Italian passages and 640 natural-language queries spanning three document styles representative of real Italian information retrieval scenarios:

• Encyclopedic passages (Wikipedia-style)
• FAQ passages (public-administration question–answer pairs)
• Legal/regulatory article excerpts (Italian legislative style)

The dataset was generated synthetically with a fixed random seed (42) using Italian-language topic vocabularies drawn from AI/NLP, legal, and public-administration domains, ensuring full reproducibility without relying on crawled or licensed data.

---

Dataset Structure

Configurations

Config	File	Rows	Description
corpus	data/corpus.jsonl	3,200	Passages to retrieve from
queries	data/queries.jsonl	640	Italian natural-language queries
qrels	data/qrels.jsonl	1,246	Query–document relevance judgements

Corpus Composition

Type	Count	Style description
wiki	1,200	Encyclopedic passages about AI, NLP, and Italian regulatory topics
faq	800	Public-administration FAQ: question + structured answer
legal	1,200	Synthetic Italian legal articles (Art. N format)

Total passages: 3,200 | Total queries: 640 | Total relevance pairs: 1,246

Field Schema

corpus.jsonl

{
  "_id":  "wiki_it_0",
  "text": "Il documento esamina intelligenza artificiale nel contesto di edilizia ...",
  "type": "wiki"
}

queries.jsonl

{
  "_id":  "q_it_0",
  "text": "Quali sono i risultati dell'uso di intelligenza artificiale nel settore sanità?"
}

qrels.jsonl

{
  "query_id":  "q_it_0",
  "corpus_id": "wiki_it_42",
  "score":     1
}

Relevance scores are binary (0 / 1). Each query has between 1 and 3 relevant documents.

---

Dataset Creation

Motivation

Existing Italian retrieval benchmarks are scarce and often require special licensing. IT-RAG-Bench was created to provide a freely available, reproducible Italian evaluation set for comparing embedding models across typical Italian enterprise retrieval scenarios (administrative portals, legal databases, public FAQs).

Generation Process

The corpus is generated from parameterised templates filled with Italian-language vocabulary lists:

• Topics: intelligenza artificiale, reti neurali, recupero dell'informazione, elaborazione del linguaggio naturale, contratti digitali, normativa GDPR, diritto del lavoro, appalti pubblici, tutela ambientale, previdenza sociale, proprietà intellettuale, riforma fiscale
• Sectors: pubblica amministrazione, sanità, istruzione, finanza, edilizia, agricoltura, trasporti, commercio elettronico, sicurezza informatica, energia rinnovabile

Queries are generated from six Italian question templates (e.g., "Quali sono i risultati dell'uso di {topic} nel settore {sector}?"). Relevance labels are assigned by randomly associating each query with 1–3 documents of a randomly selected type.

Random seed: 42 (fully deterministic and reproducible)

Caveats

Because relevance labels are randomly assigned rather than annotated by humans, absolute metric scores are lower than on human-annotated benchmarks. The dataset is best used for comparative evaluation (ranking models against each other) rather than measuring absolute retrieval performance.

---

Benchmark Results

The following results were obtained in the companion paper using FAISS HNSW indexing and nDCG@10 as the primary metric (640 queries):

Model	Type	Recall@1	Recall@5	Recall@10	MRR	nDCG@10
GE2 (Google Embeddings 2)	API	0.061	0.288	0.476	0.259	0.282
mE5 (multilingual-e5-large)	Open	0.051	0.280	0.489	0.243	0.279
E5-large	Open	0.053	0.279	0.439	0.247	0.262
mpnet	Open	0.054	0.238	0.397	0.240	0.243
BGE-M3	Open	0.046	0.253	0.404	0.224	0.238
LaBSE	Open	0.048	0.190	0.315	0.184	0.189

Experimental Setup

• Index: FAISS HNSW (M=32, efConstruction=200, efSearch=128)
• Embedding cache: SHA-256 keyed disk cache (diskcache), preventing redundant API calls
• Reproducibility: fixed seed 42, deterministic CUDA operations
• Hardware: NVIDIA A100 40GB (open-source models), API calls for commercial models

---

Usage

from datasets import load_dataset

# Load corpus, queries, and relevance judgements
corpus  = load_dataset("Siando/it-rag-bench", "corpus",  split="train")
queries = load_dataset("Siando/it-rag-bench", "queries", split="train")
qrels   = load_dataset("Siando/it-rag-bench", "qrels",   split="train")

# Example: build a corpus dict
corpus_dict = {row["_id"]: row["text"] for row in corpus}

Minimal retrieval evaluation example

from datasets import load_dataset

corpus  = {r["_id"]: r["text"] for r in load_dataset("Siando/it-rag-bench", "corpus",  split="train")}
queries = {r["_id"]: r["text"] for r in load_dataset("Siando/it-rag-bench", "queries", split="train")}

qrels = {}
for r in load_dataset("Siando/it-rag-bench", "qrels", split="train"):
    qrels.setdefault(r["query_id"], {})[r["corpus_id"]] = r["score"]

---

Repository & Code

The full benchmarking framework (embedding clients, FAISS retrieval, chunking ablations, plotting) is available at:

https://github.com/cciro94/GoogleEmbeddings2-benchmark

---

Citation

If you use IT-RAG-Bench in your research, please cite:

@misc{cirillo2026benchmarkinggoogleembeddings2,
    title     = {Benchmarking Google Embeddings 2 against Open-Source Models for Multilingual Dense Retrieval and RAG Systems},
    author    = {Stefano Cirillo and Domenico Desiato and Giuseppe Polese and Giandomenico Solimando},
    year      = {2026},
    eprint    = {2605.23618},
    archivePrefix = {arXiv},
    primaryClass  = {cs.CL},
    url       = {https://arxiv.org/abs/2605.23618}
}

---

License

This dataset is released under the Creative Commons Attribution-NonCommercial 4.0 International (CC BY-NC 4.0) license.
You are free to use and adapt it for non-commercial purposes with proper attribution.
See https://creativecommons.org/licenses/by-nc/4.0/ for details.