RetrievalRouter: Joint Modality and Architecture Selection for Document Retrieval

Hugging Face Daily Papers 08/26/26, 12:00 AM Papers

Summary

RetrievalRouter is a lightweight query-aware router that adaptively selects retrieval pipelines to improve accuracy and speed in document retrieval, outperforming static baselines.

Document retrieval increasingly supports high-stakes information access in finance, healthcare, and law. Modern retrieval pipelines vary both in modality (text or multimodal) and in retrieval architecture (dense or late-interaction). These choices impose a hard compromise: the most effective pipelines are too slow and expensive to run at scale, while the fastest fail to retrieve evidence from complex documents. Practitioners must therefore choose between missed evidence and unusable latency, with no principled basis for adapting that choice at the query level. We show that this compromise is unnecessary. Not every query requires the same pipeline. Across benchmarks spanning financial and scientific corpora, no static pipeline dominates. We introduce RetrievalRouter, a lightweight query-aware router that learns, from the query text alone, which retrieval pipeline best fits each query. A single tunable parameter exposes the full accuracy-latency frontier, and for every static baseline, RetrievalRouter offers an operating point that is simultaneously more accurate and faster. Against the best static baseline, RetrievalRouter is 2.5% more accurate and 12.4 times faster. Furthermore, compared with prior adaptive strategy selection methods, RetrievalRouter achieves significantly higher nDCG@5 across accuracy-oriented settings, while matching or numerically outperforming them on both nDCG@5 and latency in latency-oriented settings. Our code and data are available at https://github.com/emrekuruu/retrieval-router.

Original Article

View Cached Full Text

Cached at: 08/27/26, 11:19 AM

Paper page - RetrievalRouter: Joint Modality and Architecture Selection for Document Retrieval

Source: https://huggingface.co/papers/2608.25625 https://huggingface.co/login?next=%2Fpapers%2F2608.25625-

Abstract

RetrievalRouter adaptively selects retrieval pipelines per query to improve both accuracy and speed across diverse document benchmarks.

Document retrieval increasingly supports high-stakes information access in finance, healthcare, and law. Modern retrieval pipelines vary both in modality (text or multimodal) and in retrieval architecture (dense orlate-interaction). These choices impose a hard compromise: the most effective pipelines are too slow and expensive to run at scale, while the fastest fail to retrieve evidence from complex documents. Practitioners must therefore choose between missed evidence and unusable latency, with no principled basis for adapting that choice at the query level. We show that this compromise is unnecessary. Not every query requires the same pipeline. Across benchmarks spanning financial and scientific corpora, no static pipeline dominates. We introduceRetrievalRouter, a lightweight query-aware router that learns, from the query text alone, which retrieval pipeline best fits each query. A single tunable parameter exposes the fullaccuracy-latency frontier, and for every static baseline,RetrievalRouteroffers an operating point that is simultaneously more accurate and faster. Against the best static baseline,RetrievalRouteris 2.5% more accurate and 12.4 times faster. Furthermore, compared with prior adaptive strategy selection methods,RetrievalRouterachieves significantly highernDCG@5across accuracy-oriented settings, while matching or numerically outperforming them on bothnDCG@5and latency in latency-oriented settings. Our code and data are available at https://github.com/emrekuruu/retrieval-router.

View arXiv page View PDF GitHub0 Add to collection