Reranking is a second-stage retrieval step in which an initial, computationally cheap set of candidate documents or passages is reordered by a more expensive, higher-precision model that scores each candidate’s relevance to the query more accurately. It is a standard component of retrieval-augmented generation and search pipelines, where a fast retriever (such as a bi-encoder or lexical index) first narrows a large corpus down to a manageable candidate set, and a cross-encoder or learned ranker then refines the ordering. Reranking improves precision at the cost of additional latency, so candidate set sizes are tuned to balance quality and speed.