Reranking
Reranking is a second-stage retrieval step that reorders an initial set of candidate documents by relevance to the query, usually with a more precise but costlier model. It refines the fast first-pass results so the most relevant passages rise to the top before being passed to a generator.
Definition
RerankingRetrieval pipelines commonly use a two-stage design. A fast retriever — often embedding-based vector search — returns many candidates cheaply, but its ranking is approximate. A reranker then scores each query-document pair more carefully, frequently with a cross-encoder that reads the query and document together, and reorders them.
Reranking improves the precision of what reaches the model in retrieval-augmented generation, which directly affects answer quality and grounding. The trade-off is added latency and compute, so rerankers are typically applied only to a shortlist rather than the whole corpus.
Frequently asked questions
What is Reranking?
Reranking is a second-stage retrieval step that reorders an initial set of candidate documents by relevance to the query, usually with a more precise but costlier model. It refines the fast first-pass results so the most relevant passages rise to the top before being passed to a generator.
Sources
Related terms
← Full AI & prompt engineering glossary
Put Reranking to work with Prompeteer, the Agentic Contextual AI Platform →