Find the law, then state it.

§ Hybrid search

A question is answered only as well as the law it surfaces. Keyword search alone misses the opinion that says the same thing in different words; pure semantic search drifts off the exact term of art that controls. So we run both at once and fuse them. Your query is embedded once and sent down two paths in parallel — a full-text index over the literal language and a vector index over meaning — and the two rankings are merged by Reciprocal Rank Fusion, which rewards the authorities that both methods agree are on point.

Opinions and statutes are searched together off that single embedding, and the query is written to be index-friendly — ranking happens inside the database's own indexes rather than by scanning every row — so a search over the whole corpus returns in well under a second. Controlling statutes are then surfaced above the opinions that construe them, mirroring the hierarchy of authority: you start from the binding rule, not a stray case that may since have been superseded.

Embeddings

A legal-domain vector model · the query embedded once, reused for both corpora

Indexes

HNSW (approximate nearest-neighbor) for meaning · GIN full-text for keyword

Fusion

Reciprocal Rank Fusion across both rankings

Order

Controlling statutes first; opinions construe them

§ The doctrine layer

Retrieval hands the models only what the search actually found — never their training memory — and asks two distinct things. First, for each opinion returned, a fast pass distills the court's rule to a single sentence, so a results list reads as holdings rather than ranked snippets. Then a second, more capable pass synthesizes the doctrine: a plain-language statement of the governing rule, assembled only from the statutes and opinions in front of it and led by the controlling statute, because a code section binds where an older case may merely persuade.

The answer streams in as it is written, one clause at a time, each clause already carrying its citations — shown as pending until the firewall settles every badge (covered in full on the next page). Because the same question over the same sources should produce the same answer, each synthesis is cached against an exact-source key: a repeat question returns the settled, already-checked answer instantly, with no model call at all.

Holdings

A fast model pass — one rule per opinion; each case's headline holding is also pre-computed once, in advance, and served instantly

Doctrine

A more capable pass — segmented, per-clause citations, streamed as written

Grounding

Only the retrieved sources; the model is told never to add a citation from memory

Caching

Exact-source key — a repeated question skips synthesis entirely

§ Currency & supersession

Even a real, on-point opinion can quietly go stale: the Legislature amends the statute it construed, and the case now interprets text that no longer governs. We surface that automatically. The amendment history embedded in each code section's source note is parsed into a year, and when a controlling statute was amended after a cited opinion was decided, that opinion is badged as possibly predating the current text.

The check is deliberately deterministic — a date comparison, not a model judgment — so it cannot itself hallucinate. The model reads the law; arithmetic decides whether it is current. The same signal renders as “current through {year}” on the statute itself, so you can see at a glance how recently the governing text was touched.