MM-BizRAG beats SOTA by 32% on enterprise Q&A with structure-aware parsing

Most multimodal RAG systems treat all documents as images, losing the rich structure of enterprise documents like reports and slide decks. MM-BizRAG, accepted at ACL 2026 Industry Track, takes a different approach: it proactively extracts document structure through a structure-aware split that routes documents into orientation-specific ingestion pipelines. For vertically structured documents (e.g., reports), it applies explicit layout-aware parsing; for horizontally structured ones (e.g., slide decks), it uses holistic page-level representations. A unified LLM-driven artifact transformation pipeline with placeholder-based positional alignment preserves natural reading order, while inference-time multimodal assembly decouples retrieval representations from generation context—enabling richer, more grounded answers without any finetuning requirement.

Tested on a large, heterogeneous enterprise dataset and two public benchmarks (SlideVQA and FinRAGBench-V), MM-BizRAG consistently outperforms state-of-the-art vision-centric baselines by up to 32% points, with especially strong gains on report-style layouts. The authors also introduce FastRAGEval, a single-call LLM Judge metric for fine-grained generative recall that halves RAGChecker's cost while achieving stronger human alignment. This work demonstrates that explicit structure handling is key for enterprise-grade document Q&A, and the no-finetuning requirement makes it practical for real-world deployment.