Reference: Contract Extraction Pipeline

Architectural notes

Reference architecture for a pipeline that ingests contracts (PDF/DOCX),
extracts key clauses (party, term, indemnity, renewal), and stores
structured fields with provenance pointers back to source spans.

Why these picks:
- Gemini for the long-context ingest pass (whole-contract reasoning).
- Sonnet for the structured-extraction pass with strict JSON schemas.
- Modal for burst PDF processing without provisioned servers.
- Eval gates on contract holdout set prevent silent schema drift.

Common gotchas:
- PDF text extraction quality dominates accuracy; budget for OCR fallback.
- Always store source-span citations alongside extracted fields for audit.