Skip to content

README.md

Latest commit

 

History

History
131 lines (95 loc) · 4.65 KB

README.md

File metadata and controls

131 lines (95 loc) · 4.65 KB
Raw

GitLaw ⚖️

Source-grounded legal AI across 5,936 German federal laws.

GitLaw turns an unstructured legal question into something a person can inspect: relevant source paragraphs, a bounded answer, visible uncertainty and a practical next step.

Try the recruiter-first live proof →

In 20 seconds

question
   ↓
exact + BM25 + semantic retrieval
   ↓
paragraph / citation graph
   ↓
deterministic source resolution
   ↓
structured answer + uncertainty
   ↓
human-reviewable next step

The key idea is not “AI writes legal text.” It is that retrieval, evidence and verification remain inspectable separately from generation.

What it proves

  • hybrid retrieval across the German federal-law corpus;
  • paragraph-level source links and cross-reference structure;
  • deterministic citation resolution;
  • structured outputs with visible uncertainty;
  • APIs and MCP tools for agent workflows;
  • human-review boundaries for consequential legal work;
  • regression tests that turn discovered failures into permanent checks.

Proof at a glance

Signal Current repository claim
Federal laws indexed 5,936
Paragraph / graph nodes 94,178
Cross-references 200,464
FAISS vectors 98,367
Citation-resolution eval 53 / 53

Important: 53/53 measures citation-resolution regression cases, not complete legal-answer accuracy.

Public proof vs. deeper product work

The public page is intentionally simplified for a recruiter or hiring manager: question → sources → answer preview → why it matters → technical depth.

The repository contains deeper work around research workflows, Mietrecht, law-firm pilot boundaries, authenticated/professional surfaces, MCP/API capabilities and release assurance. Those layers are engineering evidence, not claims that a synthetic public demo equals production legal software.

Release assurance

GitLaw does not use “build succeeded” as a synonym for “product works.” Different surfaces have different proof levels:

Surface / risk Automated proof
Public / demo workflows Browser and routing regressions for the user-facing research flow
Mobile demo Regression checks for narrow layouts
Law-firm pilot Edge-case suite + assistant-console HTTP E2E + document-ground-truth, portal-navigation and agent-capability contracts
Pilot privacy / authority Fail-closed tests for identifiers, secrets, productive access, execution tools, consent, professional secrecy and final lawyer authority
Citation integrity Citation-resolution regression preserved in pilot CI and MCP evals
Viewer TypeScript/Vite production build + cross-domain routing regressions
MCP server Dedicated MCP CI plus optional deployment; missing deployment credentials must skip cleanly rather than produce a false-red release

The law-firm pilot edge cases deliberately include malformed German CSV, duplicate case IDs, unsupported office-file uploads, missing reviewers, non-lawyer final reviewers, missing consent/approval gates, PII/secrets, consequential action requests, excessive retention, incomplete/duplicate reviews, unsupported claims, broken citations and unverified ROI baselines.

What is not proven

  • complete legal-answer accuracy across German law;
  • production reliability for every law-firm system/vendor;
  • that a synthetic public demo is equivalent to a real law-firm pilot;
  • autonomous consequential legal action — GitLaw intentionally keeps that outside model authority.

Engineering choices

German federal-law corpus
          ↓
ingest + normalize
          ↓
BM25 / exact / FAISS
          ↓
hybrid ranker
          ↓
paragraph graph lookup
          ↓
local citation verifier
          ↓
React UI + APIs + MCP

Stack

Python · FastAPI · React · TypeScript · BM25 · FAISS · embeddings · MCP · Pydantic · Zod · CI evals

Run the key proofs

# Core/package tests
npm test

# Law-firm pilot edge cases
node --test pilot/law-firm/core.test.mjs

# Viewer build
npm install --prefix viewer
npm run build --prefix viewer

# MCP citation regression
python -m gitlaw_mcp.tests.test_eval

Run the MCP demo

python -m gitlaw_mcp.demo

Boundary

GitLaw assists research and preparation. It does not replace qualified legal advice or make consequential legal decisions autonomously.

The next meaningful proof is broader evaluation with legal professionals on anonymised matters — not simply adding more features.

Solo-built by Michael Ninh in Berlin. · AGPL-3.0