The flagship slice (docs/discovery/05): when a free-text query runs and embeddings
are available, fuse lexical top-50 (ts_rank_cd) with vector top-50 (pgvector cosine
over the structured-filtered set) using Reciprocal Rank Fusion. Placed BEFORE the
fuzzy fallback so queries with ZERO lexical hits still get semantic recall — found
and fixed during review (the fuzzy path would otherwise swallow pure-semantic
queries). Exact lexical hits keep winning (rank in both lists); deeper pages fall
through to lexical paging; ANY failure (Ollama down, nothing embedded) silently
degrades to the lexical/fuzzy path — AI can never break search.
LiveDb test proves the recall win: a 'banana' query surfaces an email with no
keyword overlap, ranked first, against real pgvector. 5 LiveDb + full suite green;
format clean.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
cesnimda
cesnimda