Reranker
A second-stage model that re-scores retrieved candidates to improve final ranking quality in RAG.
A reranker is a model that takes the top-N candidates from a fast retriever (usually vector search or BM25) and re-scores them using a more expensive but more accurate model — typically a cross-encoder. The output is a re-ordered, higher-quality shortlist for the LLM to consume.
Rerankers are the single highest-ROI addition to a RAG pipeline. Retrieval alone routinely returns 50–70% relevant chunks; reranking pushes that to 85–95% at modest additional cost.
Popular rerankers include Cohere Rerank, BGE-reranker, and cross-encoder models on Hugging Face. The trade-off: reranking N candidates costs N model calls, so N is typically kept between 20 and 100.
Key points
- Sits between fast retrieval and the LLM
- Cross-encoders are more accurate but more expensive than dense retrieval
- Typical N: retrieve top-100, rerank to top-5–10 for the LLM
- Cohere Rerank and BGE-reranker are strong off-the-shelf choices
Common use cases
Frequently asked
Related terms
A pattern that grounds LLM answers in your private documents by retrieving relevant snippets at query time.
A database optimized for storing and searching high-dimensional embeddings by similarity.
A numerical vector that captures the semantic meaning of a piece of content.
Applying Reranker in your business?
We help enterprises design, ship, and govern AI systems end-to-end.