Reranker
A reranker re-scores a short list of search results by reading the question and each result together, which is slower but more accurate than embeddings alone.
Search with embeddings is fast because documents are turned into vectors once and compared by distance, but it can miss subtle matches because the question and the document never see each other. A reranker, usually a cross-encoder, reads the question and one candidate together and outputs a relevance score. The usual recipe has two steps: an embedding model fetches the top 50 or 100 candidates, then a reranker sorts them.
Rerankers add delay to every query, because each candidate needs its own model pass, so keep the candidate list short, though listwise rerankers such as Jina Reranker v3.5 read the whole list in one pass. A reranker is a separate model from the embedding model and has its own license, so check it separately. The BGE-M3 model card, for instance, points to BAAI's bge-reranker models for re-scoring the top results.
Related terms
- Embedding model An embedding model turns text, images or other data into a list of numbers that captures its meaning, so that similar items end up close together.
- Retrieval-augmented generation (RAG) Retrieval-augmented generation lets a model answer from your own documents by retrieving the relevant passages and adding them to the prompt.
- Matryoshka embeddings Matryoshka embeddings are trained so that the first part of the vector still works on its own, which lets you cut the vector shorter to save storage and search time.