ClauseGuard is an AI-powered legal contract review platform. It helps someone without legal training understand a contract, clause by clause, in plain language.
It combines three things: a fine-tuned clause classifier, a deterministic risk engine, and an LLM (Large Language Model) that generates reports using RAG (Retrieval-Augmented Generation) — meaning it retrieves relevant passages from the document before writing an answer, instead of relying only on its training data. Every generated report cites the specific passages it used.
ClauseGuard is not a lawyer replacement. It is an explainability layer. Every risk flag traces back to a fixed rule. Every AI report traces back to a specific clause and the context retrieved for it. Every classification carries a calibrated confidence score, not a bare label.
Every clause classified and color-coded by risk the moment analysis finishes. Full walkthrough in docs/GALLERY.md.
- Upload a contract as a PDF. ClauseGuard parses it and segments it into clauses. Each clause is anchored to its exact position on the page.
- Classify every clause into one of 41 categories, derived from CUAD (the Contract Understanding Atticus Dataset). A fine-tuned Legal-BERT model does the classification. Each result carries a calibrated confidence score, with a low-confidence flag when appropriate.
- Assess risk for each clause using a deterministic rule table. No LLM is involved in risk scoring. This keeps risk levels reproducible and auditable — the same input always produces the same risk level.
- Generate a report for any clause, on demand. Each report includes a plain-language summary, the legal effect, concrete risks, and specific negotiation points. The report is grounded in supporting context retrieved from the rest of the document.
- Chat about the contract through an LLM agent. The agent can look up clauses and generate reports on its own. For anything expensive, like a full-document analysis, it pauses and asks for confirmation before proceeding.
See docs/PIPELINE.md for the full clause-analysis
flow, docs/AGENT.md for the chat/agent layer,
docs/ML.md for the classifier,
docs/DEPLOYMENT.md for the CI/CD pipeline and
self-hosted infrastructure, and docs/GALLERY.md for a
full screenshot walkthrough.
- Backend — FastAPI. Routes stay thin; a service layer owns all business logic. The LLM and embedding-model vendors sit behind a provider abstraction, so switching vendors is a config change, not a code change.
- ML — Legal-BERT, fine-tuned on CUAD, classifies each clause into one of 41 categories. Classification runs locally. A separate embedding model handles retrieval.
- Agent — A LangGraph tool-calling loop, running on Groq. Real interrupt/resume human-in-the-loop confirmation gates any consequential action.
- Retrieval — A ChromaDB vector index, scoped to each session.
- Frontend — React and TypeScript. A desktop-first document workspace: a PDF viewer with clause overlays, a docked chat sidebar, and a report panel.
- Deployment — Self-hosted, using Docker Compose. GitHub Actions runs CI/CD: it builds multi-platform images, pushes them to GHCR (GitHub Container Registry), then deploys over a Tailscale-authenticated connection. A separate n8n workflow, on its own AWS host, monitors the live application independently.
Full breakdown in docs/ARCHITECTURE.md and
docs/DEPLOYMENT.md.
| Layer | Tools |
|---|---|
| Backend | FastAPI, Pydantic, uv |
| ML | Transformers, PyTorch (Legal-BERT), scikit-learn, sentence-transformers |
| Agent / LLM | LangChain, LangGraph, Groq |
| Retrieval | ChromaDB |
| Frontend | React, TypeScript, Vite |
| PyMuPDF (backend parsing), jsPDF (client-side report export) | |
| Deployment | Docker, Docker Compose, GitHub Actions, Cloudflare Tunnel, Tailscale |
| Observability | LangSmith (LLM tracing), Umami (frontend analytics), n8n (uptime monitoring and alerting) |
Actively developed. Classifier is on its second iteration (macro-F1 0.78,
up from 0.63) after a root-cause fix to a training-label conflict — see
docs/ML.md for that writeup. This repository is a public,
documentation-focused mirror; application source lives in a private
companion repository.
These are not built yet. They are ideas under consideration for a later version.
- OCR support. Currently ClauseGuard only reads PDFs with a real text layer. OCR (Optical Character Recognition) would let it read scanned or image-only contracts too.
- Multi-language support. The classifier and prompts are English-only today. Supporting other languages would need a multilingual classifier and translated prompts.
- Clause comparison. Compare the same clause type across two contract versions — for example, redlining a revised NDA against the original.
- Whole-document summary. A single executive summary of the entire contract, generated on top of the existing per-clause reports.
- AI-assisted layout segmentation. An optional mode using a vision-capable LLM to help segment clauses in documents with unusual or inconsistent layouts, where the current rule-based segmentation struggles.
- More export formats. Reports currently export as PDF only. DOCX (Microsoft Word) export is a natural next format.
- Configurable risk rules. Let an organization define its own risk rules and thresholds, instead of using only the built-in rule table.
See LICENSE.