Search reranking usually follows an initial retrieval step that quickly gathers candidate documents. A reranker then scores the query-document pairs more carefully and sorts the candidates so the most useful results appear first. This can improve retrieval quality without running an expensive model across the entire corpus.
Imagine retrieving 50 candidate passages quickly, then having a reranker score them and keep the best five to put in a prompt. That keeps the model's context small and focused. Reranking helps most when the first-stage results contain the right answer but not near the top. It is a standard step in retrieval-augmented generation pipelines.
Reranking adds computation and latency, so measure whether it improves top results enough to justify the cost. It cannot recover a relevant document that the first retrieval stage never returned. Related: Hybrid Search, Reranker Model.