An assistant that reads every file in the firm
A law firm was burning hours digging through its own archives. We built an assistant that answers questions across their contracts and precedents, with sources.
Details anonymized and adapted for confidentiality.
The problem
A firm of around thirty people. Fifteen years of archives, thousands of matters, contracts, pleadings, judgments. It is all there, somewhere, on the file server and in the document management system.
The issue is not missing information, it is finding it. A lawyer looking for how the firm handled a non-compete clause three years ago has to remember the client name, the matter, or ask whoever worked on it. We timed a few typical searches: 25 to 40 minutes to track down an internal precedent, sometimes without finding it at all. Across the whole team, that is several billable hours a week lost to folder digging.
What we built
A RAG assistant that indexes the firm’s documents and answers natural-language questions, always with the source passages cited.
-
Matter ingestion - Connected to the DMS and the file server, extracted the text (including scanned PDFs via OCR), chunked and vector-indexed. Incremental updates: a newly filed matter is searchable the next day.
-
Search by meaning, not keywords - You can ask “have we ever argued wrongful termination of a fixed-term contract in hospitality” and get the relevant matters, even if nobody ever wrote those exact words.
-
Sourced answers, never invented - Every answer links back to the exact excerpts and the original document, clickable. If the assistant has nothing reliable, it says so instead of padding. No hallucinated case law.
-
Access partitioning - DMS permissions are respected: a lawyer only sees the matters they are entitled to. Everything stays hosted in Europe, nothing leaks to an uncontrolled public model.
The result
- Finding a precedent: from 30 minutes to under 2. The lawyer asks, reads the excerpts, opens the source document.
- Several billable hours a week recovered across the team, time spent on substance rather than searching.
- The firm’s knowledge no longer lives in one person’s head. A departed partner’s matters stay usable.
- Faster onboarding for new hires: they query fifteen years of practice instead of interrupting seniors.
- Zero invented answers observed during acceptance testing, thanks to systematic sourcing.
The stack
- Embeddings and vector search for meaning, reranking for relevance
- LLM in strict RAG mode: it only answers from retrieved documents
- OCR on scans so nothing stays out of the index
- Hosting and processing in Europe, DMS access rights enforced
- Automatic incremental sync, no full reindex