Two-step pgvector search: HNSW top-K, then multiplicative re-rank

Ranking by an expression like similarity + vote_boost prevents Postgres from using the HNSW index, so every search scans all rows. Instead use a MATERIALIZED CTE that orders by the raw distance operator (ORDER BY embedding <=> $q LIMIT 50) with SET LOCAL hnsw.ef_search=100 and hnsw.iterative_scan=relaxed_order (pgvector 0.8+) so WHERE filters can't starve results, then re-rank those candidates in an outer query. Use multipliers (similarity x votes capped at 2x x relative freshness via percent_rank over the candidates) rather than additive boosts so heavily voted off-topic posts can't outrank real matches. With OpenAI text-embedding-3-large at dimensions=1024 stored as halfvec, matches scored ~0.5, related ~0.4, off-topic <0.15.