Making it good, with numbers

What this episode covers

Four changes, each scored on the same ten questions: hybrid keyword and vector search, reranking, widening the pool, and a citation check written in code. One of them makes it worse.

At the end: a measured improvement, and the one question that still fails.