Contract PDF Extraction for Japanese Business Documents

Structure-aware contract ingestion for Japanese business PDFs, with native-text-first parsing and OCR fallback.

What this repository is

contract-ingest is a review-oriented extraction pipeline for turning contract PDFs into structured outputs with evidence references.

It is designed for:

• contract ingestion support,
• legal/business review support,
• internal document automation,
• PDF-to-JSON / CSV preprocessing.

Current strengths

• Clause/section recovery that is more robust than naive page-text parsing.
• Native text first, with OCR fallback for scanned or mixed-quality PDFs.
• Review-aware field extraction (the pipeline surfaces ambiguity instead of forcing overconfident outputs).
• Evidence-friendly outputs (evidence_refs, issue details) for downstream audit and triage.

Extracted field coverage

Current normalized extraction includes:

• contract type
• counterparties
• effective date
• expiration date
• auto-renewal
• termination notice period
• governing law
• jurisdiction

Review philosophy

This project prioritizes safe reviewability over optimistic automation.

Typical review signals include:

• missing required fields,
• low-quality field candidates,
• low confidence,
• unstable clause split,
• section boundary uncertainty.

Baseline evaluation foundation

Baseline artifacts are included for reproducible evaluation setup:

• manifest schema: data/baseline_manifest.jsonl
• baseline evaluator: scripts/eval_baseline_contracts.py
• metric definitions: docs/evaluation/baseline_metrics.md
• failure taxonomy: docs/evaluation/failure_taxonomy.md

Run baseline manifest validation:

source .venv/bin/activate
python scripts/eval_baseline_contracts.py --manifest data/baseline_manifest.jsonl

Current limitations

• Document coverage is strongest for Japanese business contracts; broader document classes may require additional tuning.
• Ambiguous cases still require human review.
• Public, adjudicated benchmark coverage is still limited (foundation is present; dataset growth is ongoing).
• The project is actively evolving; behavior can improve as layout and extraction heuristics are refined.

Quick start

uv sync
uv run pytest

Demo

Stack

Python, Pydantic, PyMuPDF, NumPy, PaddleOCR, PaddlePaddle

License

This repository is source-available for personal study, research, and evaluation.
Commercial use requires prior written permission and a separate paid license.
See LICENSE for details.