law-search.impetik.com lets a lawyer ask a natural-language question and get back an answer grounded in actual statutes — with citations, paragraph numbers, and links to the source document. Building it taught us more about retrieval engineering than any of the prior AI projects we've shipped.
Retrieval beats reasoning
We spent two months tuning the LLM. We spent six months tuning the retriever. The biggest accuracy gains came from hybrid search — sparse plus dense, with a learned re-ranker on top. The model on the end of the pipeline matters less than the chunks you feed it.
Hallucinations and how we killed them
Citations as a hard constraint. The system refuses to answer unless every claim in the response can be traced back to a retrieved chunk above a confidence threshold. The cost is occasionally saying 'I don't know' when a softer system would guess. In legal work, that's the right tradeoff.
