VLD-RAG: Agentic Framework for Multimodal Document QA
A new AI framework called VLD-RAG addresses question answering over visually-rich multi-page documents. The system, detailed in a preprint on arXiv, uses an agentic multimodal retrieval-augmented generation approach to handle evidence spread across pages with text, tables, charts, and figures. VLD-RAG builds a page-preserving multimodal index combining parsed text, metadata, and dense visual representations. It employs hybrid retrieval with keyword-based sparse search and dense semantic queries. A verifier-guided workflow coordinates three agents: Retrieval Agent, Answer Agent, and Validation Agent to broaden evidence and cross-page reasoning. The work studies multimodal RAG for long documents where retrieval must select evidence pages using both textual and visual signals.
Key facts
- VLD-RAG is an agentic multimodal RAG framework for multi-page evidence retrieval and cross-page reasoning.
- It builds a page-preserving multimodal index storing parsed text, page-level metadata, and dense visual representations.
- Hybrid retrieval combines keyword-based sparse search with dense semantic queries.
- A verifier-guided agent workflow coordinates Retrieval, Answer, and Validation Agents.
- The framework targets visually-rich documents like reports, slides, and manuals.
- Evidence needed to answer a question is often distributed across multiple pages.
- Retrieval must select evidence pages that include both textual and visual signals.
- The paper is available on arXiv with ID 2607.24748.
Entities
Institutions
- arXiv