Skip to content

ClauseGuard

Multi-agent contract & policy review pipeline

Upload a contract, get back a risk verdict — so a human reviewer starts from a report instead of a blank document.

Source on GitHub How it works

  • 3

    Specialised agents

  • 23

    Test suite

  • 2

    LLM backends

01The problem

Legal, procurement and compliance teams read long contracts by hand to spot auto-renewal traps, low liability caps and one-sided indemnity. It's slow, inconsistent between reviewers, and doesn't scale.

02The approach

Three LangGraph agents automate the first pass: the Extractor segments the document into labeled clauses, the Risk Analyzer checks each against a configurable rule set, and the Summarizer writes a plain-English verdict.

03Architecture — the camera follows the data

How it works

> Every arrow here is a Pydantic-validated hand-off.

  1. Step 1: Upload

    PDF or text

  2. Step 2: Extractor

    labeled clauses

  3. Step 3: Risk Analyzer

    rules × BM25 + embeddings

    Where the work happens

  4. Step 4: Summarizer

    plain-English summary

  5. Step 5: Verdict

    e.g. HIGH_RISK

    Where the work happens

04Perks — why it's built this way

The edge

  • Validated hand-offs

    Every hand-off between agents is checked against a Pydantic schema, so a malformed step fails loudly instead of flowing downstream.

  • Hybrid retrieval that degrades

    BM25 keyword search fused with local Ollama embeddings by reciprocal rank fusion — and keyword-only, automatically, when no embedding model is present.

  • Two backends, chosen per run

    Local Ollama for free offline runs, the Anthropic API for higher-quality paid ones. One LLM seam, so tests mock exactly one thing.

  • Evals, not vibes

    An eval harness scores the model on labeled documents — recall, precision, latency, tokens and cost per case — separate from the 23 offline unit tests.

05Tech — and what each part does

The stack

Python
Pipeline and rules
LangGraph
Three-agent state graph
FastAPI
POST /review, history, web UI
Pydantic
Schema on every hand-off
Ollama
qwen2.5:7b + nomic-embed-text, local
Anthropic API
Paid, higher-quality backend
SQLite
Review audit history
pytest
23 tests, every LLM call mocked

07Known limits — stated plainly

Where it doesn't win

  • No retry yet when the model returns malformed JSON.
  • No OCR — scanned PDFs aren't readable.
  • No human-review queue for low-confidence extractions.

→ Next file · F-04 /TBFX

Next file

TableFoxAn agent pointed at a database doesn't know the schema. Pasting the whole dump into every prompt is the usual fix — this is the better one.Decrypt it