Menu
Concepts

Reranker models

3 min readbeginnerUpdated 28 Sept 2026
1 · In one line

A reranker sorts existing text candidates by semantic relevance to a specified query.

1 · What it is

The first stage retrieves a large set of documents. A BERT reranker can score each query-passage pair for relevance. The passages are then re-ranked according to that probability.

A cross-encoder takes a query and a document as a single concatenated input and produces a relevance score. The passages are re-ranked according to that probability.

BEIR evaluated reranking the top 100 BM25 results with a cross-encoder. This second pass can improve ordering, but it has higher latency and cost than bi-encoder retrieval.

2 · Why it exists

Reranker models improve the order of candidates returned by a first-stage search.

Rescore candidatesA BERT reranker can score each query-passage pair for relevance.
Reorder resultsSemantic reranking sorts existing results by semantic relevance to the query.
Limit expensive workAzure semantic ranker reranks only the top 50 results from the initial result set.
3 · How it works

Follow a query and four retrieved candidates into a new ranking.

The reranker changes the order of an existing candidate set; it does not create that set.
  1. 1 · retrieveUse a first-stage search to obtain a candidate set.
  2. 2 · pairCombine the query with each candidate for scoring.
  3. 3 · scoreProduce one relevance score for every query-candidate pair.
  4. 4 · sortRe-rank the candidates according to the relevance probability.

Pairwise scoring is more expensive than the first retrieval pass.

4 · Where it's used
WhoWhat they askWhat it works with
Search engineer“Which retrieved passages best match this query?”Semantic reranking
RAG developer“Which candidates should enter the generator's context?”Top reranked passages
Evaluation team“Does reranking improve the order of the first-stage results?”Relevance ranking
5 · What it solves, and what it doesn't
solves
  • Rerankers sort text inputs by semantic relevance to a query.
  • Cross-encoders can evaluate a query and document together.
  • Reranking can be applied after keyword, semantic, or hybrid retrieval.
doesn't solve
  • A reranker does not retrieve the initial candidate set.
  • Azure semantic ranker cannot add information that was absent from the original result.
  • Cross-encoder reranking has higher latency and cost than a bi-encoder.
6 · Go deeper
RerankingSemantic searchHybrid searchBM25RAG

Sources used

This explainer is written in original language. The links below support its factual claims.

  1. paperPassage Re-ranking with BERT, Nogueira and Cho · read 28 Sept 2026
  2. officialReranking with Cohere models, Cohere · read 28 Sept 2026
  3. officialSemantic reranking, Elastic · read 28 Sept 2026
  4. officialSemantic ranking in Azure AI Search, Microsoft · read 28 Sept 2026
  5. paperBEIR - A Heterogeneous Benchmark for Zero-shot Evaluation of Information Retrieval Models, Thakur et al. · read 28 Sept 2026
Suggest a correction