Retrieval‑augmented generation (RAG) pipelines usually fix the indexing and retrieval configuration during preprocessing, which is ill‑suited for expert domains where queries vary in granularity, metadata constraints, and source preferences. To address this, we introduce ORDER (Optimal Routing for Dynamic Evidence Retrieval), a framework that adapts both indexing and retrieval to each incoming query. We first perform semantic clustering over a set of questions linked to a corpus, and for each cluster we learn a chunking strategy, a metadata filter, and a reranking configuration. At inference time, queries are assigned to the appropriate pre‑built index via nearest‑centroid routing. To further improve retrieval, we propose a supervised query router (QRe) that predicts which collections are most likely to contain relevant evidence, together with a Uniform Multi‑source Sampler (UMS) that distributes the retrieval budget evenly across the selected sources. Experiments on large‑scale, heterogeneous historical archives show that conditioning both indexing and retrieval on the query consistently outperforms naive baselines and strong state‑of‑the‑art RAG systems in complex expert‑domain settings.
Review: ORDER demonstrates that query‑aware routing to specialized indexes and retrieval pipelines can effectively handle heterogeneous evidence sources, offering a practical path to boost retrieval quality in specialized domains.