VLD-RAG: Agentic AI framework improves QA across multi-page visual documents

VLD-RAG, introduced by Seonok Kim, tackles the challenge of question answering over long, visually-rich multi-page documents like reports, slides, and manuals. These documents distribute evidence across pages, mixing text with layout cues, tables, charts, and figures. The framework builds a page-preserving multimodal index that stores parsed text, page-level metadata, and dense visual representations. A hybrid retrieval strategy combines keyword-based sparse search with dense semantic queries to identify candidate sources and evidence pages. This multimodal approach ensures both textual and visual signals are considered during retrieval, addressing a key limitation of prior systems that often ignore visual context.

VLD-RAG's core innovation is a verifier-guided agent workflow that coordinates three specialized agents: a Retrieval Agent, an Answer Agent, and a Validation Agent. The Retrieval Agent broadens evidence coverage, the Answer Agent generates responses, and the Validation Agent detects missing citations and refines retrieval requests when necessary. This agentic loop improves reliability by ensuring answers are grounded across scattered evidence. Evaluated on LongDocURL and MMLongBench-Doc benchmarks, VLD-RAG outperforms previous vision-based retrieval baselines in both Top-1 and Top-5 evidence-page accuracy and end-task QA generalized accuracy. The results highlight that coordinated agent verification and multimodal hybrid retrieval are crucial for robust grounding in complex document understanding.