Definition
Reranking takes a candidate set from a first-stage retriever and assigns new scores before final selection. The first stage is usually optimized to search a large corpus quickly and return enough plausible candidates. The reranker can spend more computation on that smaller set.
A cross-encoder reranker reads the query and candidate together, which allows detailed token interactions that separate embedding search does not model. Other rerankers use late interaction, learned features, metadata rules, freshness, permissions, or a blend of signals. The scoring method should match the task rather than assuming semantic relevance is the only priority.
Rodrigo Nogueira and Kyunghyun Cho's 2019 paper applied BERT to passage reranking and reported large gains on its TREC-CAR and MS MARCO evaluations. It is an influential neural reranking example, not the origin of reranking in information retrieval.
Reranking adds latency and can only score what the first stage returned. Its evaluation should therefore separate candidate recall from ordering quality. Recall at k measures whether a useful item entered the pool; metrics such as mean reciprocal rank or normalized discounted cumulative gain measure where useful items ended up.
Distinguish it from nearby terms
Retrieval searches the corpus. Reranking reorders a bounded result set. Filtering removes candidates that fail explicit conditions. Generation consumes selected context but does not itself improve the ranking.
Check your understanding
A cross-encoder places the right document first whenever it appears in the top 100, but end-to-end answers still fail. Which first-stage metric would reveal whether the document is missing before reranking begins?