also reranker, cross-encoder
A second pass that rescores the top retrieval candidates against the question with a slower, more accurate model, and keeps only the best few for the prompt.
You have done this if
You fetched fifty candidates cheaply, reranked them, and passed only the top handful to the model.
Say it in a review
We retrieve wide and rerank narrow, so the context holds a few strong chunks instead of many weak ones.
On the AI Application map Retrieval