MM-BizRAG: Rethinking Multimodal Retrieval-Augmented Generation for General Purpose Enterprise Q&A

In three linesMM-BizRAG improves multimodal retrieval-augmented generation for complex enterprise documents. The system explicitly extracts document structure via orientation-specific ingestion pipelines (layout-aware parsing for vertical reports, holistic page representations for horizontal slide decks), then assembles multimodal contexts at inference. Up to 32% gains on SlideVQA and FinRAGBench-V without fine-tuning.

Summary generated by Claude — human-verified