What is a re-ranker in RAG? What models are used (cross-encoder, bge-reranker)?

In RAG (Retrieval-Augmented Generation), the re-ranker is a component that re-evaluates and sorts documents or fragments retrieved from a knowledge base to select the most relevant ones for answer generation.

Main types of re-ranker models:

• Cross-encoder: a model that takes a pair (question, document) and processes them simultaneously, calculating relevance. This is a more accurate but computationally expensive approach, as each document is processed together with the query.

• BGE-reranker (Bi-Encoder): models that encode the question and documents separately into vectors, then compute cosine similarity. Faster but may be less accurate than cross-encoder.

In RAG, a fast bi-encoder is often used for preliminary selection, followed by a cross-encoder for precise ranking. The re-ranker helps improve the quality of document selection, which enhances the final answer generation.