Contractex¶
Contractex is a Python library for legal document intelligence. It provides the processing layer that legal AI products are built on: chunking, extraction, retrieval, privacy enforcement and quality measurement. It is not an application.
Most serious legal tooling is closed-source or built in-house at large firms. That leaves small practices, clinics and advice organisations dependent on vendors. An open, auditable processing layer lets them build their own, and measure whether it is accurate enough to rely on.
pip install contractex
Design principles¶
| Principle | In the code | Measured by |
|---|---|---|
| The deterministic/probabilistic boundary is explicit. | Every task declares whether it calls a model; the task table says what happens when the output is poor. Structure parsing, PII detection, citation extraction, classification and playbook checks make no model calls. | Tests on the deterministic parts; the evaluation harness for the rest. |
| Every extracted value is traceable to its source. | ClauseAwareChunker returns exact substrings; ProvenanceTracker resolves a value to character offsets in the original document. | Provenance benchmark on CUAD. |
| Privacy is enforced in code at the call site. | A PrivacyProfile on each document; PrivacyAwareLLMRouter blocks, rejects or redacts on every model call made by tasks, pipelines and RAG. No provider is ever chosen for you. | Routing matrix and privacy tests. |
| Quality is measurable and gateable. | EvalHarness scores any extraction function against labelled suites; assertions fail CI when accuracy drops. | The benchmarks, regenerated by one command and checked in CI. |
What Contractex refuses to do¶
- Put the text of a
secretdocument into any model prompt. - Send a
restricteddocument to anything but a local model (LocalProvider). - Send
confidentialorrestrictedtext to a model without redacting the personal data it detects. - Choose a model vendor for you: a task that needs a model and was not given one raises an error.
- Install any model vendor's SDK unless you ask for it.
Where it stops¶
Without Presidio installed, names are not detected. Scanned PDFs need OCR. Chunking and citation patterns are English-centric. The Limitations page lists everything known.
Start here¶
- Installation: what to install for your setup.
- Quick start: parse, redact, chunk and run tasks on a sample NDA.
- Privacy model: what each sensitivity level enforces, and where.
- Benchmarks: chunking, provenance and privacy, measured.
Contractex is open source under the Apache License 2.0 and maintained by Quiet Signals Lab.