Unsupervised Query Routing for Retrieval Augmented Generation

Query routing for retrieval-augmented generation aims to assign an input query to the most suitable search engine. Existing works rely heavily on supervised datasets that require extensive manual annotation, resulting in high costs and limited scalability, as well as poor generalization to out-of-distribution scenarios. To address these challenges, we introduce a novel unsupervised method that constructs the "upper-bound" response to evaluate the quality of retrieval-augmented responses. This evaluation enables the decision of the most suitable search engine for a given query. By eliminating manual annotations, our approach can automatically process large-scale real user queries and create training data. We conduct extensive experiments across five datasets, demonstrating that our method significantly enhances scalability and generalization capabilities.