Built a private, secure AI assistant trained on 15,000+ legal documents that helps lawyers retrieve case precedents, summarise contracts, and draft responses in seconds.
Challenge
A regional law firm with 45 lawyers was losing billable hours to document research. Associates were spending 3–5 hours per case searching through internal precedent databases, reading through contracts, and cross-referencing regulatory guidance. Institutional knowledge was siloed — when senior partners retired or left, their expertise walked out with them. The firm had explored off-the-shelf legal AI tools but found them too generic and concerned about sending client data to third-party cloud services.
Workflow approach
We built a fully private, on-premises RAG (Retrieval-Augmented Generation) system deployed in the firm's own cloud environment. The system ingests and indexes the firm's complete document library: case files, contracts, precedents, regulatory guidance, and research memos — 15,000+ documents in total. We implemented hybrid retrieval (vector similarity + BM25 keyword search) using Weaviate as the vector store, with a custom reranking layer using a cross-encoder model fine-tuned on legal text. The LLM layer runs on a private Claude deployment via AWS Bedrock, ensuring no data leaves the firm's infrastructure. The assistant answers questions with citations, generates contract summaries, flags potential issues in documents, and drafts initial responses to common legal queries — all with the firm's specific precedents and jurisdiction as context.
Observed outcome
Lawyer research time dropped 70% on average across the firm. Junior associates now handle initial research tasks that previously required senior review, improving leverage ratios. The system's answer accuracy rate (evaluated against a labelled test set of 500 legal queries) is 94%, with clear citation of source documents. The firm estimates 2,400+ billable hours recovered in the first year, representing over $600,000 in recaptured revenue.
Outcome metrics
- Reduction in research time
- 70%
- Documents indexed
- 15,000+
- Average query response
- < 3s
- Answer accuracy rate
- 94%
Systems involved
- Claude 3.5 (AWS Bedrock)
- Weaviate
- LangChain
- Python
- FastAPI
- Docker