Reranker
Step 2 of a RAG pipeline — after the vector DB fetches ~50 candidates, a cross-encoder re-orders them by more detailed relevance. Return the top 5 to the LLM.
Vector search is "bi-encoder" (query and document encoded separately ⇒ fast). The reranker is "cross-encoder" (query + document encoded together ⇒ accurate but 100× slower). The two-stage design balances speed and precision. Tools: Cohere Rerank (managed), BGE Reranker (open weights, BAAI), Jina Reranker, voyage-rerank, sentence-transformers cross-encoders. Along with hybrid search, the single biggest quality lift in modern RAG systems.