A reranker double-checks the initial search results and puts the best ones first.
DefinitionWhat it means
A reranker is a model, typically a cross-encoder, applied after an initial retrieval step to re-score a shortlist of candidate results against the original query more precisely than the fast vector search that produced them. Because rerankers compare the full query and document together rather than relying on pre-computed vectors, they are more accurate but too slow to run over an entire dataset, so they only touch the top handful of candidates.
Why it mattersWhy you should care
Rerankers are the standard second stage in production RAG and search pipelines because raw vector similarity alone often surfaces results that are topically related but not actually the best answer to the query. Adding a reranking step is one of the highest-leverage, lowest-effort improvements teams make when a retrieval system's answers feel close but not quite right.
At a glanceSee it
Opening the reranker box — a cross-encoder scores each candidate one at a time by letting query and document tokens attend jointly, then sorts by score.
The candidate-depth decision and its failure mode — a reranker can only reorder what retrieval already fetched, so a document outside the top k is unrecoverable.
Where you see itIn the wild
- A reranking stage added after initial vector search in a RAG pipeline
- Cross-encoder models like those from Cohere used to rerank results
- A/B testing showing reranking improves answer relevance