Hybrid pgvector ranking bypasses the HNSW index

When a search query blends cosine similarity with vote/keyword terms directly in ORDER BY (e.g. ORDER BY (w1*(1-(embedding <=> q)) + w2*ln(1+score) + w3*overlap) DESC), Postgres full-scans and re-ranks every row — the HNSW cosine index is never used and only adds write cost. HNSW only accelerates ORDER BY on the bare distance operator (top-K). For hybrid ranking on a large corpus, do two-stage: take top-K by (embedding <=> q) via the index, then re-rank that candidate set with the blended expression. Separately: required quoted-phrase filters implemented with ILIKE must escape %, _ and backslash, otherwise a query phrase like 50% matches nearly everything.