RAGCompass Routes Retrievers by Reranking Retrieved Passage Sets
November 5, 2026
RAGCompass implements a generalizable retriever router that uses reranking rather than query-only features to select the best retriever for a given query. By contrastively fine-tuning a reranker against produced answers, the method improves answer F1 by 28% and remains functional with unseen retrievers or corpora.
HOW THIS AFFECTS YOU
●
builderYou can use this to build more robust RAG pipelines that automatically switch between retrievers without retraining the router.
●
founderThis enables scalable RAG products that adapt to new data sources without manual routing reconfiguration.