Reordering retrieved candidates with a stronger relevance model.
Reranking applies a more expensive comparison after a fast retriever produces candidates. It can improve precision, but it cannot recover relevant material omitted from the candidate set.