Contractex

Contractex is a Python library for legal document intelligence: the processing layer for chunking, extraction, retrieval, privacy enforcement and quality measurement that legal AI products are built on.

This project aims to democratise legal tech by building a transparent open-source library that changes who within the legal field gets to benefit from the latest developments in machine learning and AI.

pip install contractex

Documentation: quietsignalslab.com/contractex/docs

What it does

• Parses contracts deterministically: sections, defined terms, cross-references.
• Extracts parties, clauses and terms with the model you choose: a local model through Ollama, or OpenAI, Anthropic or Google.
• Traces an extracted value back to character offsets in the source.
• Enforces a privacy profile on every model call.
• Measures quality with an evaluation harness and CI gates.

What it refuses to do

• Put the text of a secret document into any model prompt.
• Send a restricted document to anything but a local model.
• Send confidential text to a model without redacting the personal data it detects.
• Pick a model vendor for you, or install one's SDK unless you ask for it.

from contractex import LegalDoc
from contractex.privacy import PrivacyBlockedError, PrivacyProfile
from contractex.tasks import TaskRegistry

doc = LegalDoc(
    full_text="Board minutes: acquire Acme Corp for $4.2bn. Contact 415-555-0132.",
    privacy_profile=PrivacyProfile(sensitivity="secret"),
)
pipeline = TaskRegistry.default().build_pipeline(
    ["pii_detection", "summarization"],
    task_kwargs={"summarization": {"llm_provider": "llama3.1:8b"}},  # a local Ollama model
)
try:
    pipeline.run(doc)
except PrivacyBlockedError:
    print("blocked before any prompt reached the model")

blocked before any prompt reached the model

Measured, not claimed

Generated by python -m benchmarks and checked in CI. Methodology and limitations are in the benchmarks documentation.

Limits

Without Presidio (contractex[privacy]), names are not detected. Scanned PDFs need OCR (contractex[ocr]). Patterns for chunking, citations and personal data are English-centric. The full list is on the Limitations page.

Project

Changelog · Contributing · Security · Apache License 2.0 · maintained by Quiet Signals Lab