DocuWeave-Bench

A retrieval benchmark for evaluating PDF chunking strategies across five document domains.

Overview

DocuWeave-Bench is a dataset of 6,100 question-answer pairs grounded in 421 real-world PDFs spanning five domains: research papers, technical documentation, legal documents, financial reports, and medical literature.

It was designed to evaluate and compare PDF chunking strategies for Retrieval-Augmented Generation (RAG) pipelines. The benchmark was used in the paper:

DocuWeave: Layout-Aware PDF Chunking for Retrieval-Augmented Generation (2026)

Dataset Statistics

Split	# PDFs	# QA Pairs	Avg Q/PDF
test	417	6,100	~14.6

Domain	# PDFs	# QA Pairs
research_papers	92	~1,340
technical_docs	87	~1,270
legal_docs	83	~1,210
financial_reports	79	~1,150
medical_docs	76	~1,130

Benchmark Results (R@1)

Chunker	R@1	R@3	R@5	NDCG@10	MRR
DocuWeave	0.2862	0.4475	0.5149	0.4336	0.3837
Naive (fixed-size)	0.1975	0.3495	0.4351	0.3958	0.3001
Recursive	0.1790	0.3307	0.4143	0.3659	0.2792
LangChain	0.1475	0.2707	0.3395	0.2938	0.2291
Semantic	0.1184	0.2018	0.2518	0.2742	0.1753
PDFPlumber	0.1126	0.1926	0.2393	0.2649	0.1660

Embeddings: BAAI/bge-base-en-v1.5. Retrieval: FAISS flat index, top-10. All differences vs DocuWeave are statistically significant (Wilcoxon, p < 0.001).

Files

data/
  qa_pairs.jsonl       # 6,100 QA pairs (one JSON object per line)
pdfs/
  research_papers/     # 92 PDFs
  technical_docs/      # 87 PDFs
  legal_docs/          # 83 PDFs
  financial_reports/   # 79 PDFs
  medical_docs/        # 76 PDFs

QA Pair Schema

Each line in data/qa_pairs.jsonl is a JSON object:

{
  "query":          "What is the authors' proposed method?",
  "answer":         "The authors propose ...",
  "question_type":  "factoid",
  "domain":         "research",
  "pdf":            "datasets/pdfs/research_papers/arxiv_2301_00001.pdf",
  "gold_chunk_id":  "a3f7b2...",
  "gold_text":      "We propose DocuWeave, a layout-aware PDF chunker ...",
  "section_title":  "3. Method",
  "page_start":     4,
  "page_end":       5
}

Usage

from datasets import load_dataset

# QA pairs
ds = load_dataset("mrjvenky18/docuweave-bench", data_files="data/qa_pairs.jsonl")

# Or download a PDF
from huggingface_hub import hf_hub_download
pdf_path = hf_hub_download(
    repo_id="mrjvenky18/docuweave-bench",
    repo_type="dataset",
    filename="pdfs/research_papers/arxiv_2301_00001.pdf"
)

Citation

@article{jannegorla2026docuweave,
  title   = {DocuWeave: Layout-Aware PDF Chunking for Retrieval-Augmented Generation},
  author  = {Jannegorla, Venkateswararao},
  journal = {arXiv preprint},
  year    = {2026},
}

License

The benchmark QA pairs and scripts are released under CC BY 4.0. PDFs are sourced from public repositories (arXiv, SEC EDGAR, PubMed, government portals) and redistributed under their respective open licenses.