What is reranking?

Reranking re-orders search results by scoring each candidate against the query with a model that reads both together. A fast retriever fetches, say, 100 candidates by vector similarity; the reranker reads each pair closely and promotes the truly relevant ones, sharply improving what the LLM finally sees.

Embedding search is fast but shallow: it compares meanings compressed into single vectors. A reranker is slow but deep: it cross-reads query and document. The standard architecture uses both, wide retrieval then precise reranking, because each covers the other’s weakness.

Rerankers are cheap for what they do; the catalog’s reranker lists at $0.10 per million tokens. In most RAG systems, adding one is the highest-leverage relevance fix available.

Related terms