ContractGuard

Japanese contract risk analysis built as an AI engineering case study — LangGraph workflow + pgvector RAG + multi-modal ingestion + recoverable streaming UX.

⚠️ Not a legal service. This repository has never been operated commercially — Japan Attorney Act §72 (弁護士法第72条) reserves paid legal advice for licensed attorneys. The codebase is published as an open-source technical artifact only. Outputs are not legal opinions.

中文文档 | 日本語 | License

Status

Production-ready open-source reference implementation. The full stack — frontend, backend, OCR, payment, email, Postgres, Redis, error tracking — is wired with real integrations and ready to deploy. It has simply never been launched, by design (Attorney Act §72).

A synthetic Japanese contract sits in docs/samples/ so the local flow can be exercised end-to-end immediately after clone.

Related MCP server: Contract Review AI MCP

Architecture

flowchart LR
  U[React/Vite UI<br/>text, PDF, image upload] --> API[FastAPI routers]
  API --> Q[Quote + PII + OCR budget guards]
  Q --> PAY[KOMOJU checkout<br/>reference implementation]
  PAY --> JOB[Persistent analysis job]
  JOB --> SSE[Recoverable SSE stream<br/>status + events + after_seq]
  JOB --> LG[LangGraph pipeline]
  LG --> P[parse_contract]
  P --> A[clause-by-clause risk analysis]
  A --> T[tool call: analyze_clause_risk]
  T --> RAG[(PostgreSQL pgvector<br/>331 Japanese legal articles)]
  A --> S[tool call: generate_suggestion<br/>medium/high risks only]
  S --> REP[report generation + translation]
  REP --> CACHE[(Redis 72h report cache)]
  REP --> DB[(PostgreSQL orders/reports/costs)]

Tech Stack

Quick Start (local)

Local run only requires an OpenAI API key.

cp .env.example .env
# Edit .env: set OPENAI_API_KEY
docker compose up --build

Then open http://localhost:5173 and upload docs/samples/sample-contract-ja.txt.

In this minimal mode:

• ✅ Plain-text contracts and text-based PDFs (selectable text) work end-to-end.

• ❌ Image / scanned-PDF OCR is disabled. To enable it, add GOOGLE_APPLICATION_CREDENTIALS_JSON and GOOGLE_VISION_PROJECT_ID.

• KOMOJU / Resend auto-bypass in dev — no real payment, no real email.

Production Setup

The repository is shaped to deploy to production by setting APP_ENV=production and supplying credentials for each external service:

When APP_ENV=production, the app refuses to boot if any of the above is missing or FRONTEND_URL still points at localhost. Strict-validation logic lives in backend/config.py (validate_runtime()).

fly.toml and vercel.json describe the deployment topology used during development. The service is not currently hosted.

Flow

1. Upload a contract (text, PDF, or image). The upload route runs text extraction, PII checks, token estimation, non-contract detection, and OCR budget guards.

2. Checkout reference path creates an order. Empty KOMOJU credentials trigger a local bypass in dev.

3. /review/:orderId starts or resumes the persistent analysis job and streams progress events that survive page refresh.

4. LangGraph parses clauses, analyzes each clause with RAG-grounded tool calls, and generates suggestions only where the risk warrants it.

5. /report/:orderId shows the saved report, clause excerpts, risk filters, and PDF export — retained for 72 hours.

User contract text is deleted after analysis. The vector store contains only public e-Gov statutes; user contracts are never embedded.

Demo

Repository Map

• backend/agent/graph.py — LangGraph pipeline.

• backend/agent/tools.py — RAG-grounded tool calls.

• backend/services/analysis_executor.py — persistent analysis job + event sourcing.

• backend/rag/store.py — pgvector storage and search.

• backend/config.py — runtime configuration and strict validation.

• frontend/src/pages/ReviewPage.tsx — recoverable analysis progress UI.

• frontend/src/pages/ReportPage.tsx — report UI with risk filters and PDF export.

• tests/ — backend pytest suites.

• scripts/smoke_local_flow.sh — end-to-end local smoke test.