Czech Labour Law Knowledge Service

English · Česky

A cited-answer RAG service over 4 Czech labour laws, usable by people (web), programs (REST API) and AI agents (MCP). Live demo: lukas-rag.vercel.app (Czech UI, English toggle top right)

What it does

• Searches 2,475 paragraphs of 4 Czech labour laws. Every result has a paragraph ID like ZP § 52 odst. 1 and a source link.
• Answers only from the retrieved text, with citations. When the law does not say, it answers “I don’t know” and marks evidence_status: insufficient.
• One shared core, three doors: web UI, REST API (FastAPI) and MCP server (official SDK). Same results everywhere.
• Measured, not guessed: held-out questions, confidence intervals, offline quality gate in CI.

Human ──────── Web UI (Vercel) ──────┐
AI agent ───── MCP server (stdio) ───┼──> tools/service.py ──> tools/rag.py (retrieval pipeline)
Programs ───── REST API (FastAPI) ───┘

Connect your agent (MCP)

git clone https://github.com/zenkovski/rag-demo.git && cd rag-demo
python -m venv .venv && .venv/bin/pip install -r requirements-service.txt
claude mcp add czech-labour-law -e RAG_OFFLINE=1 -- "$PWD/.venv/bin/python" "$PWD/tools/mcp_server.py"

Claude Desktop config, Windows paths and troubleshooting: docs/mcp.md. RAG_OFFLINE=1 = no API key needed, nothing is paid, answers come from the cache (82 prepared questions).

Tools: search_knowledge (no LLM in mode=fast), ask_knowledge (cited answer), get_document (full paragraph). Ask your agent: “Use search_knowledge: Jak dlouhá je výpovědní doba?” (“How long is the notice period?”). Real output (text shortened):

{
  "query": "Jak dlouhá je výpovědní doba?",
  "mode": "fast",
  "passages": [
    {"rank": 1, "id": "ZP § 51 odst. 2", "law": "ZP", "title": "Výpověď",
     "text": "(2) Výpovědní doba činí nejméně 2 měsíce, s výjimkou výpovědi dané a) …",
     "score": 0.8907, "source_url": "https://www.zakonyprolidi.cz/cs/2006-262"}
  ]
}

ask_knowledge adds answer, evidence_status, citations (marker, id, excerpt, source_url) and deterministic checks. Passage text is returned as untrusted data, never as instructions (tested with an injected instruction).

REST API quick start

export RAG_API_KEYS=my-local-key     # or put it in .env
docker compose up --build            # offline by default; docs at http://localhost:8000/docs

curl -s -X POST localhost:8000/api/v1/search -H "X-API-Key: my-local-key" -H "Content-Type: application/json" \
  -d '{"query": "Jak dlouhá je výpovědní doba?", "k": 3}'
curl -s -X POST localhost:8000/api/v1/ask -H "X-API-Key: my-local-key" -H "Content-Type: application/json" \
  -d '{"question": "Jak dlouhá je výpovědní doba?"}'
curl -s "localhost:8000/api/v1/documents/ZP%20%C2%A7%2051%20odst.%202" -H "X-API-Key: my-local-key"

Without Docker: pip install -r requirements-service.txt, then RAG_OFFLINE=1 RAG_API_KEYS=my-local-key uvicorn api:app --app-dir tools. Errors are structured (401, 404, 413, 422, 503 provider down, 504 timeout) and carry a request ID. On Windows, send Czech text from a UTF-8 file: --data-binary @query.json.

Key results

Measured on 134 questions the system was never tuned on (sets v4 + v5, frozen in git before the run; 112 answerable, 22 where the law has no answer). Details and caveats: docs/evaluation-v5.md.

Honest scope: in this run the LLM steps (rewrite, pick, answer) were done by Claude Sonnet with the same prompts, not by the DeepSeek models of the live site, to spend no API credits; questions and judges are also Claude, with a 15-answer spot check. Without any LLM, the fast search alone finds the gold paragraph in the top 5 for 101/112. Older measurements (66 questions, ablations with p-values) are in docs/benchmark-results.md; method in docs/evaluation-methodology.md; known failures in docs/failure-gallery.md; limits in docs/limitations.md.

How it works

question
  → rewrite into legal language            ("výplata" → "mzda")             LLM (DeepSeek v4.1 Flash)
  → semantic + keyword search              (e5-large + BM25, fused by RRF) → 20 candidates
  → LLM picks 5 passages                   (+ deterministic rules: paragraph 1, in-section references)
  → answer only from those 5, with [n] citations                            LLM (DeepSeek v4 Pro)
  → deterministic checks                   (citation range, numbers present in the cited text)

The same retrieval exists in Python, LangChain (LCEL) and JavaScript (live site); parity tests keep them identical. The service layer and both APIs call the Python version and do not copy it.

Run the checks

pip install -r requirements-service.txt -r requirements-dev.txt
python -m pytest tests                                  # 165 tests, offline, network blocked
ruff check . && mypy
python tools/experiments.py --check --no-base           # quality gate: retrieval vs results/baseline.json

CI runs the same: lint + types, tests, quality gate, and a Docker build with a smoke test.

Documentation

Engineering audit · MCP guide · Architecture · Design decisions · ADRs · Incidents · Threat model · History v1 → v3.1 · AI-assisted development. Most lab documents are in Czech.

Authorship

I designed the system and directed the build with Claude Code; I verified it with tests, evaluations and code reviews. Not legal advice: a technical demo. Law texts from zakonyprolidi.cz, downloaded 2026-10-06 (262/2006, 435/2004, 187/2006, 590/2006 Sb.).