Reranking
Reranking is a second search step that takes a short list of candidate documents and reorders them by relevance to the query, using a more accurate but slower model than the first search.
Also called: reranker, re-ranking, reranker model
By Divya Sharma · Checked Nov 15, 2026 · How we review
First-stage search, whether vector search, BM25 or hybrid search, has to be fast across millions of items, so it compares the query and each document in a cheap way. A reranker only looks at the top few dozen candidates, so it can afford to read the query and each document together and judge relevance more carefully.
Reranker models
A reranker model takes a query and a document and returns a relevance score. Most are cross-encoders: one transformer reads both texts at once. Hosted options include Cohere Rerank 5 and Voyage Rerank 3; open-weight ones include Jina Reranker v3.5. Some teams also use an LLM as the reranker, which costs more per query.
Example
A retrieval-augmented generation app retrieves a few dozen chunks for "refund window for annual plans". The reranker moves the annual-plan refund clause from deep in the list to first place, and only the top handful go into the prompt. Fewer, better chunks mean a shorter prompt and less room for the model to latch onto the wrong passage.
Limits
A reranker cannot rescue a document the first stage never retrieved, so first-stage recall still matters. It adds latency and cost per query. The retrieval and reranking guide compares options.
Related terms
- Cross-encoder
- A cross-encoder is a model that reads a query and a document together in one pass and outputs a relevance score. It is more accurate than comparing separate embeddings, but too slow to run over a whole collection.
- Hybrid search
- Hybrid search runs a keyword search and a vector search on the same query and merges the two ranked lists, so results can match on exact words as well as on meaning.
- Retrieval-augmented generation
- Retrieval-augmented generation (RAG) is a pattern where a system first searches a collection of documents for passages relevant to a question, then gives those passages to a language model so its answer draws on them.
- Late interaction (ColBERT)
- Late interaction is a retrieval method, introduced by ColBERT, that encodes the query and each document separately into one vector per token, then scores a match by comparing those token vectors in a cheap final step.
- Retrieval metrics (recall@k, nDCG, MRR)
- Retrieval metrics score how well a search system returns relevant documents. Recall@k checks how many relevant items appear in the top k, nDCG rewards ranking them high, and MRR measures how early the first correct one appears.