RetrievalRouter is a lightweight router that learns from query text alone which retrieval pipeline — varying by modality and by dense versus late-interaction architecture — best fits each incoming query, rather than forcing a single static pipeline to handle every query at a fixed accuracy-latency trade-off. Across benchmarks spanning financial and scientific document corpora, no single static pipeline dominates, and RetrievalRouter offers an operating point that is simultaneously more accurate and faster than every static baseline tested. Against the best static baseline it is 2.5% more accurate and 12.4 times faster, and the authors have released their code and data.
