Reranking
Also: re-ranking, reranker, cross-encoder reranking
A cheap, recall-oriented retriever (BM25 or a bi-encoder) pulls the top N candidates, then a cross-encoder reads each query–document pair together and rescores it, and only the best few go to the LLM. This adds cross-encoder precision at bi-encoder scale.
Example
A bi-encoder retrieves the top 50 chunks out of 1M by cosine similarity; a cross-encoder scores just those 50 query–chunk pairs and the top 3 are passed to the LLM, pushing an embedding false positive out of the final set.