A second-stage model or scorer that re-orders top retrieval candidates for relevance—often lifts RAG quality more than swapping vector DBs alone.
Bi-encoder retrieval (query embedding vs document embedding) is fast but coarse. Rerankers (cross-encoders or lightweight rankers) read query–passage pairs jointly and assign finer scores.
Trade-off
Reranking adds latency and cost—apply to the top k results, not the whole corpus.