Keyword search was never built for case law

Ask a lawyer how they research and you will usually hear a version of the same ritual: guess the keywords, run the boolean query, skim forty results, adjust the keywords, run it again. The tools have gotten prettier over the years, but the contract underneath has not changed since the 1970s: you describe the words you hope appear in the judgment, and the system finds documents containing those words.

The problem is that precedent does not live in words. It lives in fact patterns, in legal tests, in the shape of a situation. Two decisions can describe the same self-defence scenario with almost no vocabulary in common. A boolean query that catches one will miss the other, and the lawyer never finds out what they missed.

What hybrid retrieval actually does

Building CaseScout, we learned quickly that neither semantic search nor keyword search wins on its own. So we fuse them. Dense embeddings, trained specifically for legal text, capture meaning, so a question about a fact pattern surfaces decisions that discuss that pattern in any vocabulary. Classic sparse retrieval catches what embeddings fumble: exact statute sections, party names, terms of art where the literal string matters. The two result lists are fused, and a cross-encoder reranker reads the top candidates against the actual question before anything reaches the next stage.

• •Meaning first: "can police search a car after a traffic stop" finds the relevant jurisprudence even when no decision phrases it that way.
• •Literal when it counts: a query containing a neutral citation or section number treats that as a hard signal. An exact hit is never demoted by a fuzzy one.
• •Reranked, not just retrieved: pulling 500 candidates is easy; the discipline is reading them against the question and keeping the thirty that matter.

Retrieval without verification is a liability

Better search creates a new risk: a fluent AI summary that cites cases it never actually read. In law, one fabricated citation can end a career, which is why we treat verification as part of retrieval, not an afterthought. In our pipeline, the writing model can only cite documents that were actually opened during the research run, and every verbatim quote is mechanically checked against the source text before it is allowed into the memo. If the corpus does not support a proposition, the memo says so out loud instead of improvising.

The standard is not "sounds right." The standard is "traceable to a real document a lawyer can open."

Keyword search asked lawyers to think like an index. Modern retrieval lets them think like lawyers: ask the question in plain language, get an answer grounded in the actual corpus, and check every citation with one click. That is not a nicer search box. It is a different contract.