Reranker

A Reranker scores the tens or hundreds of documents returned by initial retrieval using a cross-encoder that is more sophisticated than a simple vector-similarity measure. It selects the few documents most relevant to the question and is a key component of RAG pipelines. Omitting reranking can introduce irrelevant documents into the LLM prompt, reducing answer quality or increasing hallucinations. Commercial and open-source options such as Cohere Rerank and BGE-Reranker are widely used, often together with relevance thresholds that exclude low-scoring documents entirely.

Business applications and impact#

Improves RAG answer quality by supplying the LLM with the most relevant context.

Implementation guidance and considerations#

Set relevance-score thresholds to keep irrelevant or noisy documents out of the prompt.