voilaj/swiss-legal-rag-bench

Swiss Legal RAG Bench

An end-to-end benchmark for retrieval-augmented generation systems operating on Swiss federal and cantonal law.

Modelled on Butler & Butler, Legal RAG Bench (Isaacus, March 2026 — arXiv:2603.01710), this benchmark evaluates a RAG pipeline along three orthogonal dimensions and decomposes errors into the failure-mode taxonomy that paper established.

Why a Swiss benchmark

Swiss law is multilingual (DE/FR/IT/RM), federal-cantonal layered, and under-represented in the international legal-AI evaluation literature. The OpenCaseLaw search-quality benchmark measures retrieval accuracy only — it cannot tell us whether a generative model deployed on top of our retrieval actually produces a grounded and correct answer, nor can it decompose end-to-end errors into hallucination / retrieval / reasoning components. Swiss Legal RAG Bench fills that gap.

Methodology

For each question i, embedder e, and generator LLM l, three binary signals are computed:

Signal	Definition
c_eli (correctness)	Model's answer entails the reference answer.
g_eli (groundedness)	Model's answer is supported by retrieved passages.
r_eli (retrieval accuracy)	At least one annotated supporting passage was retrieved.

Errors decompose into four leaves (Butler & Butler 2026, Figure 1):

Class	Condition
Correct	g = 1 ∧ c = 1
Hallucination	g = 0
Retrieval err.	g = 1 ∧ c = 0 ∧ r = 0
Reasoning err.	g = 1 ∧ c = 0 ∧ r = 1

This decomposition triangulates which component (embedder vs LLM) drives end-to-end error.

Dataset

10 hand-crafted federal-law questions across the three official languages and the major substantive areas (Schuldrecht, Strafrecht, Verfassungsrecht, Sachenrecht, Mietrecht, Verfahrensrecht, Verjährung, Wirtschaftsfreiheit).

{
  "id":          "q-001",
  "language":   "de",
  "legal_area": "Schuldrecht / Haftpflichtrecht",
  "difficulty": "basic",
  "question":   "Welche vier Voraussetzungen müssen kumulativ erfüllt sein, damit eine ausservertragliche Haftung nach Art. 41 OR entsteht?",
  "reference_answer": "Schaden, Widerrechtlichkeit, adäquater Kausalzusammenhang ...",
  "evidence": {
    "statutes":  [{"law_code": "OR", "sr_number": "220",
                   "article": "41", "language": "de"}],
    "decisions": ["bge_BGE_132_III_122"]
  },
  "claim_type": "elements_of_norm"
}

evidence.decisions lists canonical OpenCaseLaw decision_ids (matches the IDs in mcp.opencaselaw.ch/entscheid/<id>). evidence.statutes lists the SR-number + article that authoritatively answers the question.

Versioning

v0.1 (28 April 2026) — anchored to the OpenCaseLaw corpus snapshot of that date (969,738 decisions). Adding questions or fixing annotations bumps the patch version; changing the evaluation methodology bumps minor.

Baseline (v0.1, 2026-04-28)

Live OpenCaseLaw stack (BM25 + RRF + Haiku rerank, top-K = 5) + Claude Sonnet 4.6 generator + Sonnet judge:

Metric	Score
Correctness	100.0 % (10/10)
Groundedness	90.0 % (9/10)
Retrieval accuracy	70.0 % (7/10)

Error decomposition: 9 correct, 1 hallucination, 0 retrieval, 0 reasoning.

The single hallucination is q-001 (Art. 41 OR conditions). The generator correctly listed the four classical elements (Schaden / Widerrechtlichkeit / Verschulden / Kausalzusammenhang), but the causation element is a doctrinal addition not literally in either the Art. 41 OR text or in the retrieved passages. The judge correctly flagged it: legally correct, ungrounded. This is precisely the failure mode Butler & Butler 2026 identified.

By language:

Lang	n	Correct	Grounded	Retrieved
de	6	100.0 %	83.3 %	66.7 %
fr	3	100.0 %	100.0 %	66.7 %
it	1	100.0 %	100.0 %	100.0 %

Per-question results: see results/run_2026-04-28.json in the GitHub mirror.

Evaluation harness

The harness (evaluate.py) takes pluggable retriever / generator / judge functions:

python3 -m benchmarks.swiss_legal_rag_bench.evaluate \
    --mcp-url https://mcp.opencaselaw.ch \
    --top-k 5 \
    --output results/run.json

Defaults call the live OpenCaseLaw MCP for retrieval and Claude Sonnet 4.6 for generation + judging. Swap in any embedder, LLM, or judge to compare.

Limitations

• Initial size: 10 seed questions. This is small. The Isaacus benchmark used 100 hand-crafted questions; we will grow toward that target via expert annotation.
• Coverage: federal law only. Cantonal procedural and material law is out of scope for v0.1.
• Reference answers are concise (1–4 sentences) — by design, to keep the entailment judge reliable.
• Annotated evidence is not exhaustive. A question's evidence.decisions may list one leading case where multiple are equally authoritative; the retrieval-accuracy metric is satisfied if any listed evidence passage is retrieved.

Citation

@misc{opencaselaw_swiss_legal_rag_bench_2026,
  author       = {OpenCaseLaw contributors},
  title        = {Swiss Legal RAG Bench: an end-to-end RAG benchmark for the Swiss federal corpus},
  year         = {2026},
  howpublished = {Hugging Face Datasets},
  url          = {https://huggingface.co/datasets/voilaj/swiss-legal-rag-bench}
}

Methodological reference:

@misc{butler_legal_rag_bench_2026,
  author       = {Butler, Abdur-Rahman and Butler, Umar},
  title        = {Legal RAG Bench: an end-to-end benchmark for legal RAG},
  year         = {2026},
  eprint       = {2603.01710},
  archivePrefix= {arXiv}
}

License

CC0-1.0 (public domain dedication). Same licence as the underlying OpenCaseLaw corpus.

Links

• GitHub source: https://github.com/jonashertner/caselaw-repo-1/tree/main/benchmarks/swiss_legal_rag_bench
• Underlying corpus: https://huggingface.co/datasets/voilaj/swiss-caselaw
• Live MCP server: https://mcp.opencaselaw.ch
• Verification architecture explainer: https://opencaselaw.ch/#verification