RetrievalRouter: Joint Modality and Architecture Selection for Document Retrieval
A query-level router beat fixed retrieval pipelines on both accuracy and latency.
The paper argues that document retrieval systems should not choose one modality or architecture for every query. RetrievalRouter selects the retrieval pipeline from the query text alone, spanning text or multimodal and dense or late-interaction options. Across financial and scientific benchmarks, the authors report no static pipeline dominated. Against the best static baseline, they report 2.5% higher accuracy and 12.4x lower latency. HF Daily Papers' note
The paper argues that document retrieval systems should not choose one modality or architecture for every query. RetrievalRouter selects the retrieval pipeline from the query text alone, spanning text or multimodal and dense or late-interaction options. Across financial and scientific benchmarks, the authors report no static pipeline dominated. Against the best static baseline, they report 2.5% higher accuracy and 12.4x lower latency. HF Daily Papers' note
score 4