Pursuant to nothing but engineering honesty: retrieval quality, latency and cost, measured on a quote-anchored golden dataset and versioned in git.
Same 52 answerable questions, three retrieval configurations. Hybrid fuses pgvector cosine search with Postgres full-text via reciprocal rank fusion; reranking re-orders the hybrid top-30 with a local cross-encoder (ms-marco-MiniLM-L6-v2, CPU).
Wall-clock per query on CPU-only hardware — no GPU anywhere in the stack. The retrieval path calls no paid API: embedding and reranking are local models, so retrieval cost per query is $0. Generation cost is metered separately at the /ask endpoint.
Every run is a committed JSON file; this table is read from those artifacts.