⚖️ LexMind AI

AI legal contract review, grounded in the text it reads.

LexMind is a review assistant, not a draft or review tool. It never invents clauses, legislation, or facts. Every conclusion cites the contract text it was asked to read — see app/prompts/skills.py.

---

The Brief

A FastAPI + LangGraph pipeline that ingests a legal agreement (PDF, DOCX, or plain text), inventories its clauses, scores risk, flags missing protections, and produces an evidence-backed executive summary with negotiation recommendations — each finding traceable to the source text.

Status: MVP code scaffold. Deliberately narrow scope, production-quality foundations.

---

How It Works

1. Parse — normalize PDF / DOCX / TXT into clean text (app/parsers/)
2. Inventory — identify and catalogue the agreement's clauses
3. Assess — score risk per clause against skill-specific prompts (app/prompts/skills.py)
4. Detect — surface missing clauses a careful reviewer would expect
5. Report — executive summary + negotiation recommendations, with source_spans pointing back into the contract

flowchart LR
    A[Agreement<br/>PDF · DOCX · TXT] --> B[Parser]
    B --> C[Clause Inventory]
    C --> D[Risk Scoring]
    C --> E[Missing-Clause Detection]
    D & E --> F[Executive Summary]
    F --> G[source_spans → contract text]

Loading

---

Repository Layout

lexmind/
├── backend/
│   ├── app/
│   │   ├── main.py          # FastAPI entrypoint
│   │   ├── api/reviews.py   # POST /api/v1/reviews · GET /api/v1/health
│   │   ├── core/config.py   # settings from environment / .env
│   │   ├── llm/client.py    # mock | anthropic | openai | ollama
│   │   ├── parsers/         # pdf / docx / txt normalization
│   │   ├── prompts/skills.py # one prompt template per pipeline skill
│   │   ├── schemas/         # Pydantic output contracts (JSON schemas)
│   │   ├── workflow/        # LangGraph state, nodes, graph, service
│   │   └── scripts/run_review.py  # CLI runner
│   ├── tests/
├── examples/
│   ├── contracts/           # sample gas supply agreement
│   └── output/              # generated review JSON
├── eval/
│   └── test_contracts/      # labelled corpus for precision/recall
├── docs/
└── fly.toml                 # Fly.io deployment (syd)

---

Running

cd backend
python -m venv .venv && .venv\Scripts\activate   # Windows
pip install -r requirements.txt

# CLI (deterministic mock LLM — no API key needed)
python -m app.scripts.run_review ../examples/contracts/sample_gas_supply_agreement.txt

# API
uvicorn app.main:app --reload
# open http://127.0.0.1:8000/docs

---

Configuring a Real LLM

Copy .env.example to .env and set LLM_PROVIDER:

Provider	Notes
anthropic	Default model claude-sonnet-4-20250514
openai	—
ollama	Local, offline

The pipeline is provider-agnostic; the mock is deterministic and offline, so the test suite and CLI run with no API key.

---

Tests

cd backend
pytest -q

---

Design Notes

• Traceable evidence — every risk conclusion carries source_spans pointing back into the contract text (PRD AI principles)
• JSON contracts — all output models live in app/schemas/, shared by the pipeline and the API
• Composable graph — the workflow is a compiled LangGraph in app/workflow/graph.py; nodes are pure functions in nodes.py, individually testable and replaceable

---

Out of Scope (v1)

Contract drafting, Word redlining, multi-user collaboration, version comparison, e-signatures, billing, workflow automation, and case-law retrieval. See the PRD for the full roadmap.