From 3-hour document searches to 3-second answers
A London legal firm with 8 researchers was losing 2–3 hours per person daily to manual document search across 50,000+ files. We built a RAG system that lets them query the entire library in plain English with full citations in under 3 seconds.
Eight legal researchers were spending 2–3 hours daily searching through a document library of over 50,000 files.
The search tool was keyword-based with no semantic understanding. Junior researchers were consistently missing relevant precedents.
RAG was the right architecture because the problem was retrieval, not generation. We ingested 50,000+ documents with custom chunking logic that respects legal document structure.
Embeddings were generated using a model fine-tuned on legal text with Pinecone configured with rich metadata for filtered retrieval.
Week 1: Data audit and ingestion pipeline. Week 2: Embedding generation (38 hours processing time). Week 3: Retrieval testing — 200 benchmark queries, initial accuracy 87%. Week 4: Accuracy reached 98%, UI integration. Week 5: Full rollout to all 8 researchers.
Each researcher saves 20+ hours per week. Complex briefs reduced from 3 days to 4 hours. Research quality standardised upward — every researcher now has access to the same comprehensive retrieval that previously required senior partner knowledge.
BY THE NUMBERS
The difference is like going from a library with no index to one where you can ask the librarian anything in plain English.
Discovery calls are free and no-pitch. We'll tell you honestly if we can deliver results like this for you.