ModelRefs / Reranking — AI Glossary
Reranking — AI Glossary
A post-retrieval step using a cross-encoder or LLM judge to score and reorder retrieved documents by relevance to the query.
Overview
Reranking models (Cohere Rerank, BGE Reranker, FlashRank, Jina Reranker) take (query, document) pairs and output a relevance score. Unlike bi-encoders, cross-encoders perform full attention between query and document, enabling much higher precision at the cost of O(k) model calls. Typical pipeline: retrieve top-100 with ANN, rerank to top-5 for generation.
Reference details
| Topic | rag |
|---|---|
| Also known as | cross-encoder reranking, neural reranking |
| Last reviewed | 2026-06-24 |
Related terms
Example: Where it sits in the pipeline
Retrieve → rerank → generate. Retrieval optimises recall and returns perhaps 50–100 candidates; reranking optimises precision and cuts that to the 3–5 that fit the prompt; generation sees only the survivors. Skipping the middle stage means the model reads whatever the ANN index happened to rank highest, which is a coarser signal than it looks.
Commonly confused with
Reranking is the stage; a reranker is the model that performs it. The stage can also be filled by an LLM judge or a heuristic such as recency weighting or reciprocal rank fusion — a cross-encoder is the common choice, not the only one.
When to use it
Reach for it when:
- The generation step has room for far fewer passages than retrieval returns
- You fuse several retrieval strategies and need one coherent ordering
- Groundedness metrics show the right passage present but ranked too low
Reach for something else when:
- Retrieval already returns few, well-ordered candidates
- Every millisecond counts and precision is adequate
- It is being used to compensate for chunking that should be fixed upstream
Referenced by
This term is used by the following ModelRefs references:
Where this appears on ModelRefs
Continue your research
Use these connected ModelRefs sections to compare alternatives, inspect implementation paths, and review the evidence and governance boundaries relevant to Reranking — AI Glossary.
Frequently asked questions
What is Reranking?
A post-retrieval step using a cross-encoder or LLM judge to score and reorder retrieved documents by relevance to the query.
Is Reranking the same as cross-encoder reranking?
Yes — cross-encoder reranking, neural reranking are common aliases for Reranking.
What concepts are related to Reranking?
Closely related concepts include dense retrieval, late interaction, bi encoder, context recall.