Optimizing Retriever Selection for LLMs via Learning to Rank

Retrieval-Augmented Generation (RAG) systems commonly depend on a single, static retriever, even though prior work suggests that no individual retriever is universally effective across all query categories. In this work, we investigate a query routing strategy that adaptively chooses among multiple retrievers based on the input query, leveraging both heuristic, train-free methods and learned routing models. We formulate retriever selection as a learning-to-rank (LTR) task and propose LTRR, a framework that learns to rank retrievers according to their expected contribution to downstream LLM performance. Through experiments on diverse question-answering benchmarks with controlled variations in query types, we demonstrate that routing-based RAG consistently surpasses the strongest single-retriever baselines. The gains are particularly substantial when training with the Answer Correctness (AC) objective and when using pairwise ranking methods, with XGBoost yielding the best results. Additionally, our approach exhibits stronger generalization to out-of-distribution queries. Overall, our results underscore the critical role of both training strategy and metric choice in effective query routing for RAG systems.