Back to glossary

Term

Reranking

Reranking reorders an already-retrieved list of hits with a more accurate model — typically a cross-encoder that scores query and each candidate together, instead of just comparing vector distances.

Reranking — explained

Embedding-based search is fast but coarse: each vector is generated independently, and the model has no idea at indexing time which question will be asked later. Cross-encoders, in contrast, see query and candidate together and can judge more finely which hit actually answers the question — but they are too slow to run over millions of documents. The standard solution: first retrieve 50–100 candidates via vector or hybrid search, then narrow them down to the top 5–10 with a reranker.

Example / Practical context

A RAG chatbot pulls 50 chunks from the vector database and sends them to a reranker model like Cohere Rerank, BGE-Reranker (open source), or Voyage rerank-2. The model rescores every (query, chunk) pair; the top 5 are forwarded to the LLM. Answer quality and reduction of irrelevant context tokens improve measurably — at the cost of extra latency (typically +100–300 ms) and API fees.

Reranking is not retrieval but a second stage after it (two-stage retrieval). Hybrid search instead works at the same level before reranking — it combines two retrievers but does not re-judge content. Reciprocal Rank Fusion is a specific fusion method inside hybrid search and not a full reranker.

Entdecke mehr

Themenuebersicht