Case 04 · Law Firm Tri-Corpus RAG (Cases + Contracts + Statutes) · Clause Search 45min → 3min · Citation Accuracy >95%¶

TL;DR: Law Firm Tri-Corpus RAG (Cases + Contracts + Statutes) · Clause Search 45min → 3min · Citation Accuracy >95%. Direction: RAG Knowledge Base. Scale: mid-size Chinese law firm. Hard KPI: clause search P50 45min → 3min, citation accuracy >95%. Tech: bge-large + BGE-reranker, tri-corpus tiered retrieval, version-aware routing. My role: RAG architecture, tri-corpus tiering, embedding fine-tuning, lawyer co-design. Client name anonymized under NDA.

• Direction: RAG Knowledge Base
• Industry: legal
• Scale: mid-size Chinese law firm
• Hard KPI: clause search P50 45min → 3min, citation accuracy >95%

Problem¶

Lawyer clause search takes 45 minutes on average, versions are messy, citation errors hurt output quality, and the three corpora (cases / contracts / statutes) are siloed.

Solution¶

bge-large + BGE-reranker, tri-corpus tiered retrieval, clause-level citation provenance, version-aware routing, client contracts on separate tenant.

Tech Stack¶

bge-large + BGE-reranker, tri-corpus tiered retrieval, version-aware routing

Result (Hard KPIs)¶

clause search P50 45min → 3min, citation accuracy >95%

My Role¶

RAG architecture, tri-corpus tiering, embedding fine-tuning, lawyer co-design

Related¶

• Full 20-case portfolio · Direction · RAG Knowledge Base
• PDF one-pager
• 中文原页

Next: Download PDF · 30-min alignment

Client name anonymized · KPIs verified · Real client name disclosable under NDA.