Reranker models

A second-stage neural model (cross-encoder) that re-ranks initial retrieval results to maximize relevance and precision for final output.

Reranker models are critical for high-performance Retrieval-Augmented Generation (RAG) and semantic search. They operate in a two-stage pipeline: a fast first-stage retriever (like a bi-encoder or vector search) quickly pulls a large candidate set, perhaps 100 documents. The reranker, typically a Transformer-based cross-encoder, then processes each query-document pair together. This deep, contextual analysis yields a precise relevance score, a capability bi-encoders lack. The result is a highly refined ranking, ensuring the Large Language Model (LLM) or end-user receives the most accurate context. This refinement directly translates to better metrics: some benchmarks show a 35% reduction in LLM hallucination and significant gains in NDCG@10 scores.

https://github.com/AnswerDotAI/rerankers