In short. Retrieval reranking orders retrieved passages by relevance to a query before selected passages are supplied to a language model.

Definition: Retrieval reranking scores passages already returned by a search system against a query and reorders them by relevance. In retrieval-augmented generation, a process that supplies retrieved material to a language model, reranking helps select the passages passed to the model before it writes an answer.

Key facts

The implementation details below reflect documentation reviewed on 17 October 2026.

Cohere's Rerank overview describes ordering supplied documents from most to least semantically relevant to a query.
The Sentence Transformers retrieve and rerank guide separates candidate retrieval from a subsequent cross-encoder, a model that processes a query and candidate text together to score relevance.
Cohere's Rerank API reference provides top_n to limit the number of returned results.
Microsoft's semantic ranker documentation describes reranking an existing result set, with only the top 50 results proceeding to semantic ranking.

How does reranking differ from retrieval?

Retrieval creates the candidate set. Reranking evaluates candidates against the question again. The Sentence Transformers pipeline guide allows the retrieval stage to use lexical search, which matches query words, or dense retrieval, which compares numerical representations of meaning.

Its cross-encoder then processes each query and candidate document together. This joint processing provides a relevance score for ordering the candidates. The guide explains the reason for the initial retrieval stage: scoring a large document collection as query-document pairs is slow. A smaller candidate set makes that subsequent scoring practical.

This creates a dependency between the stages. A passage excluded from the candidate set remains outside that reranking operation, as Microsoft's semantic ranker documentation makes explicit for its implementation.

Where does it fit before answer generation?

Microsoft's RAG overview describes semantic ranking as optional in its classic retrieval-augmented generation pattern. In that pattern, application code passes search results to a language model, which formulates an answer from them. The following schematic shows the sequence when reranking is included:

Query + retrieved passages
    -> Reranker scores and orders candidate passages
    -> Application selects passages for the answer context
    -> Language model generates an answer using that context

Cohere's API reference documents controls for the selection step: top_n limits returned results, while max_tokens_per_doc limits the text processed per document through truncation. These settings affect how much material survives into later processing. An implementation review therefore examines both the candidate passages and the text actually supplied to the reranker.

How should a website team evaluate the results?

Cohere's best practices describe scores as query dependent. Its threshold-setting procedure uses representative queries paired with borderline relevant documents, then uses their average score as a reference for filtering. This makes evaluation specific to the intended questions and document collection.

For content preparation, Microsoft's semantic ranker documentation identifies information-rich prose as suitable input, including knowledge bases and online documentation. For a website team, that supports reviewing the actual passage containing an answer when assessing retrieval results. The documented unit being scored is the supplied text relative to a query.

FAQ

Can you rerank keyword search results?

Yes. The Sentence Transformers guide explicitly allows lexical search to supply candidates to its cross-encoder reranker.

Can reranking recover a passage missing from the candidates?

Microsoft's semantic ranker documentation limits that operation to existing search results. If a needed passage is absent, the retrieval stage needs attention.

How can you choose a relevance threshold?

Cohere's best practices recommend testing representative queries with documents judged borderline relevant for the intended use. Their scores provide a reference for choosing a cutoff.

Sources

Check how a page reads to answer engines. Open the GEO score analyzer