Reranker

A reranker is a second-stage retrieval component that takes an initial candidate set (typically the top-k results from a faster first-stage retriever) and produces a refined ranking using a more accurate but more expensive scoring model. Rerankers are typically cross-encoders — transformer models that take both the query and a candidate document together and predict relevance — whereas first-stage retrievers use bi-encoders that compute query and document embeddings independently. Common rerankers include Cohere Rerank (production-grade managed service), BGE-Reranker (open source), Jina Reranker, mxbai-rerank, and various Cross-Encoder models from the Sentence-Transformers library. Rerankers can dramatically improve RAG answer quality, especially when the first-stage retriever returns many candidates that are loosely related and need fine-grained relevance ranking. The trade-off is latency and cost — rerankers add 50-200ms and per-document inference cost, multiplying with top-k. AI governance teams document reranker choice and threshold configuration in their RAG pipeline lineage. Production architectures typically use 10-50 first-stage candidates reranked to top-3 or top-5 for the final LLM context.

Reranker integration with Centralpoint: Centralpoint supports rerankers from Cohere, BGE, Jina, and other providers in its model-agnostic RAG pipeline, with per-skill configuration of candidate count and threshold. Tokens are metered across rerank and generation, prompts stay local, and reranked chatbots embed through one line of JavaScript on any portal.