Legal-clause RAG with citation

Problem

Drafting attorneys needed to find precedent clauses across 12 years of past deals. Search was full-text only — too noisy, missed semantically similar wording.

Users

Internal — 40-lawyer corporate practice

Key flow

1. 1Lawyer enters clause type or natural-language description
2. 2System returns ranked precedents with side-by-side comparison
3. 3Lawyer copies clause into draft, source is recorded

What I built

• Hybrid BM25 + dense retrieval over 8k contracts split into clauses
• Pinpoint citation — answer links back to exact paragraph and document
• Confidence-weighted ranking with deal-type and jurisdiction filters
• Audit log — every search and result archived for compliance

Limits

• Confidential matters are tagged and never returned to other practice groups
• No drafting — surfaces precedent but doesn't generate clauses