Reranking

Reranking is a second scoring pass over what retrieval fetched. The first pass is built for speed across everything stored: embedding similarity narrows millions of chunks to the twenty or fifty that look closest. The reranker then reads the question and each candidate together, as one input, and scores how well this text answers this question. That comparison is more precise than embedding similarity but slower, so it runs only on the small candidate list.

The two passes measure different things. Embedding similarity says two texts are near each other in meaning, and by that measure a passage about the question can outscore the passage containing the answer. The reranker scores the narrower property, whether the passage answers the query, and only the best few go to the model.