Research / Project detail

RAG & trustworthy knowledge services

Document understanding, retrieval, and trustworthy generation, from industrial OCR robustness to budgeted multi-document evidence allocation.

RAGDocument AIEvidence reasoning
Problem

Retrieval and OCR quality alone do not guarantee reliable answers. How can systems allocate evidence across documents and account for document noise?

Method

Combining retrieval, document structure, evidence allocation, and generation evaluation. Related papers cover OCR robustness for RAG, budgeted multi-document evidence allocation, and visual-document RAG.

Outcomes

InduOCRBench spans 12 industries, 570 documents, 3,402 pages, and 11 challenge types. BEAR evaluates hierarchical retrieval and multi-document reasoning under a shared generator and evidence budget.

Trustworthy generation framework
Trustworthy generation framework
Results & evaluation

Results & evaluation

InduOCRBench · Industrial-document robustness

Visual-style documents reach 82.9% OCR accuracy but 52.8% RAG answer accuracy, a 30.1-point gap. Table 2 compares the Avg scores of ten methods on general and industrial-document benchmarks.

InduOCRBench · Industrial-document robustness
OCR methodOmniDocBench Avg ↑InduOCRBench Avg ↑
PP-StructureV386.760.3
MinerU275.566.5
Doc2x84.661.6
GPT-4o75.052.0
Qwen3-VL-235B89.270.9
Gemini-2.5 Pro88.074.5
Deepseek-OCR87.061.5
Hunyuan-OCR94.168.1
MinerU2.590.772.5
PaddleOCR-VL92.978.2

When Good OCR Is Not Enough · Table 2 / Figure 2 ↗

InduOCRBench: end-to-end evaluation of OCR errors, retrieval, and generation
InduOCRBench: end-to-end evaluation of OCR errors, retrieval, and generation · When Good OCR Is Not Enough · Figure 2 ↗

BEAR · Budgeted evidence allocation

DragonBall retrieval comparisons use a Qwen3-32B generator, with evidence contexts aligned to approximately 4K tokens where applicable; LongRefiner uses a different auxiliary model. Values are mean ± standard deviation over three runs. BEAR(docs) achieves 34.63 ± 2.17 EM and 48.02 ± 1.45 F1 on 2Wiki.

BEAR · Budgeted evidence allocation
MethodRecall % ↑Completeness % ↑Hallucination % ↓Irrelevance % ↓
BM2529.17 ± 1.0635.13 ± 1.8432.55 ± 2.4032.25 ± 1.95
BGE-M358.17 ± 1.9159.56 ± 2.4528.81 ± 2.2611.54 ± 1.55
RAPTOR31.93 ± 1.6659.78 ± 2.7026.50 ± 2.4813.67 ± 1.29
LongRefiner29.24 ± 2.3243.86 ± 1.9834.88 ± 2.2521.26 ± 1.87
HippoRAG268.80 ± 1.7573.37 ± 2.2419.55 ± 1.846.93 ± 0.93
BEAR(docs)85.55 ± 0.4791.86 ± 1.325.64 ± 1.022.40 ± 0.49
BEAR(nodes)84.25 ± 0.8488.64 ± 1.527.60 ± 1.123.71 ± 0.86

BEAR · Table 1 / Figure 2 ↗

BEAR: offline hierarchical indexing and online coarse-to-fine exploration and recovery
BEAR: offline hierarchical indexing and online coarse-to-fine exploration and recovery · BEAR · Figure 2 ↗
Related publications

Related papers

Related coverage

Related coverage