What is reranking?
Reranking is a step after first-pass retrieval in which a finer model re-evaluates how relevant each candidate is to the query and sorts the candidates by score. It usually handles only the top candidates, because comparing items one by one costs more than the first-pass retrieval.
Coarse filter first, then fine ranking
The Sentence-Transformers documentation splits the process into two stages. The first uses keyword search or vector retrieval to quickly pull a shortlist, for example 100 candidates, out of a large collection of documents. The second uses a cross-encoder that reads the query and each candidate together and outputs a relevance score from 0 to 1.
The second stage is usually more accurate but slower, so running it over millions of records one by one is impractical; it is applied only to the first stage's candidate list. Cohere's Rerank is a service of this kind, reordering documents by semantic relevance.
Where it fits in tool retrieval
When an AI agent has many tools available, a system can first retrieve a batch of tools with BM25 or semantic retrieval, then use a reranking model to pick the few that best fit the task and hand them to the model to choose from. A tool that does not stay on the reranked list is invisible to the model. In the terms of TSO, this is the candidate exposure stage.
Sources and scope
The source-language entry records a source check on 2026-11-04. This English edition does not represent a new check of the cited pages.