Hemlock Legal's associates were spending 6+ hours per week searching through case law, contracts, and internal memos to answer routine questions. The knowledge was scattered across SharePoint, email, and a legacy document management system with no unified search.
Previous attempts at enterprise search had failed because keyword matching couldn't handle the nuance of legal language. Associates needed answers that cited specific clauses, not just document links. And the system had to know when it didn't know, hallucinated legal advice could be catastrophic.
We built a retrieval-augmented generation system using LangChain and Pinecone for vector search over 1,400 indexed documents. The assistant answers questions in plain language, cites specific sources with page numbers, and explicitly flags low-confidence responses rather than guessing.
The ingestion pipeline handles PDFs, Word documents, and emails with automatic chunking optimized for legal document structure. We built a custom evaluation framework with 500 question-answer pairs reviewed by senior partners to measure accuracy before launch. The UI is a React chat interface embedded in their existing intranet.
The assistant achieved 94% accuracy on the firm eval set, reviewed and validated by senior partners. Associates report saving an average of 6.2 hours per week on research tasks. The system has processed over 12,000 queries in its first three months with zero reported hallucination incidents.
Our associates trust it, and they're lawyers, so that's saying something. It cites its sources, and when it's not sure, it says so. That's exactly what we needed.