SEC-RAG-Bench

A hand-written benchmark of 100 questions for evaluating RAG (Retrieval Augmented Generation) systems over SEC 10-K filings. Every question is anchored to a specific passage in a real filing, making it suitable for measuring Recall@K, Precision@K, Citation Accuracy, and refusal behavior.

Dataset summary

• 100 questions across 6 categories
• 8 companies: Apple, Google, Microsoft, Nvidia, Tesla, Costco, Netflix, Amazon
• Fiscal years 2022-2024
• Ground truth by construction: each question anchored to a real passage with company, year, and exact snippet recorded

Categories

Category	Count	What it tests
Factual	32	Single-chunk fact retrieval
Numerical	17	Exact-match numbers and dollar amounts
Multi-hop	14	Information spanning two filings
Comparison	8	Cross-filing or cross-year comparisons
Negation	14	"Not", "excludes", precision-sensitive language
Unanswerable	15	Model should refuse rather than fabricate

Dataset structure

JSONL format, one question per line. Example:

{
  "id": "q_001",
  "category": "factual",
  "question": "What is the address of Apple's principal executive offices?",
  "answer": "One Apple Park Way, Cupertino, California 95014",
  "source_company": "apple",
  "source_year": 2023,
  "source_snippet": "One Apple Park Way, Cupertino, California 95014"
}

Fields

• id: unique question identifier (q_001 through q_100)
• category: one of factual, numerical, multi_hop, comparison, negation, unanswerable
• question: the natural-language question
• answer: the expected answer (or null for unanswerable)
• source_company: company whose filing contains the answer (lowercase)
• source_year: fiscal year of the source filing (integer), or "multiple" for cross-year questions
• source_snippet: exact text passage containing the answer (null for unanswerable)

Usage

Load with 🤗 Datasets

from datasets import load_dataset

ds = load_dataset("YOUR_USERNAME/sec-rag-bench")
print(ds["train"][0])

Load with pure Python

import json

questions = []
with open("benchmark_v1.jsonl") as f:
    for line in f:
        questions.append(json.loads(line))

Methodology

Each question was authored by hand against real 10-K passages. A paraphrase rule was enforced: question wording cannot trivially overlap with the source chunk, so retrieval has to work on meaning, not keyword matching.

Six categories were chosen to stress different retrieval failure modes:

• Factual tests the easy case (single chunk, direct answer).
• Numerical tests exact-match retrieval (dollar amounts, line items).
• Multi-hop tests cross-filing synthesis (same company, two years).
• Comparison tests cross-entity synthesis (two companies, same year).
• Negation tests precision on words like "not", "excludes".
• Unanswerable tests refusal behavior: the model should say "I don't know".

Evaluation metrics

Recommended metrics for using this benchmark:

• Recall@K: did the correct source chunk appear in the top-K?
• Precision@K: what fraction of top-K retrieved chunks were relevant?
• Citation Accuracy: did the generated answer cite the correct source number?
• Faithfulness (RAGAS): is every claim in the answer supported by context?
• Answer Relevancy (RAGAS): does the answer address the question?
• Refusal Accuracy: for unanswerable questions, did the model correctly refuse?

Known limitations

• Strict single-source labeling: SEC filings often publish the same information in multiple places (comparative-year data, cross-sections). The current ground truth credits only one source per question. The real recall rate is higher than measured when alternative sources are accepted.
• Comparison category is small (n=8): per-category numbers for comparison questions are noisy.
• English only, US filings only: no coverage of international filings or non-English language.

Source project

The benchmark was built as part of SEC-RAG-Bench, an evaluation-driven RAG platform. The full project includes hybrid retrieval (vector + BM25 via RRF), semantic caching, a multi-provider LLM gateway, and Prometheus/Grafana observability.

Citation

If you use this benchmark in research or writing:

@misc{sec-rag-bench-2026,
  author = {Neha Rani},
  title = {SEC-RAG-Bench: A Benchmark for Retrieval-Augmented Generation over SEC Filings},
  year = {2026},
  publisher = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/YOUR_USERNAME/sec-rag-bench}}
}

License

MIT

Contact

Built by Neha Rani. Issues and improvements welcome via the GitHub repo.