The framing that vector search finds candidates while reranking decides relevance is the mental model that fixed my retrieval quality. Hybrid dense plus lexical mainly buys you recall on exact terms and rare tokens that embeddings smear together, and the reranker is what stops that wider net from dumping noise into the prompt. I'd add that context compression only pays off after reranking is solid, otherwise you're compressing documents that shouldn't have made the cut.