Reranking reorders a small set of candidate results with a more precise model, often a model that reads the question and each passage together. More costly than the initial search, it is applied only to the best candidates to improve final relevance.
Example
Keep the 50 best passages from the search, then re-rank them and keep the top 5.