We argue that multi-document reasoning is constrained not only by how much text a model can read, but also by how a limited query-time evidence budget is allocated across documents and semantic granularities. Full-context inference exposes the model to broad evidence non-selectively and at high per-query cost, while flat chunk retrieval often returns locally relevant passages that are weakly organized for cross-document synthesis. We present BEAR, a framework for structured evidence allocation that builds hierarchical semantic indices offline and performs coarse-to-fine evidence access at query time through complementary exploration and recovery paths. This design formulates coarse-to-fine reasoning as structured evidence allocation under a fixed final evidence-context budget. Across nine evaluation settings spanning synthetic evidence allocation, multi-hop QA, agentic retrieval, conversational memory, long-document QA, and open-domain QA, BEAR delivers strong coverage–faithfulness trade-offs on DragonBall, achieves the best retrieval-based results on 2WikiMultiHopQA, LoCoMo, QASPER, MuSiQue, and PopQA, and remains competitive on HotpotQA and NarrativeQA. Under the matched retrieval protocol, BEAR operates with an approximately 4K-token final evidence context, substantially smaller than the inputs used by the long-context references. These results show that evidence exposure—not context length alone—should be selectively organized, with gains arising from hierarchy as an allocation substrate and complementary exploration and recovery rather than semantic chunking alone.

Industrial Retrieval-Augmented Generation (RAG) systems depend on optical character recognition (OCR) to transform visual documents into text. Existing OCR benchmarks rely on character-level metrics, which inadequately measure downstream RAG effectiveness under real-world conditions. We introduce an OCR benchmark for industrial RAG systems covering 11 challenging document types, including extreme layouts, high-resolution pages, complex or watermarked backgrounds, historical documents with non-standard reading orders, visually decorated text, and documents containing tables and mathematical formulas. Evaluating recent SOTA OCR models under a controlled OCR-first RAG pipeline shows clear performance degradation on realistic industrial documents despite strong conventional benchmark scores. We find that high OCR accuracy does not necessarily translate into strong downstream RAG performance: structural and semantic errors can cause substantial retrieval failures even when WER/CER remains low. Further analysis shows that this mismatch is category-dependent, arises through both retrieval-side and downstream generation-side failures, and remains stable across representative OCR-first pipeline choices. The benchmark is publicly available at https://github.com/Qihoo360/InduOCRBench.