MAGE-RAG: Multigranular Adaptive Graph Evidence for Agentic Multimodal RAG in Long-Document QA

2026-06-14 · Source: Artificial Intelligence · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics · Depth: Expert, quick

Summary

MAGE-RAG is a multigranular adaptive graph evidence framework designed for agentic multimodal RAG in long-document question answering. It addresses limitations of existing RAG methods that struggle with locating sparse evidence across text, tables, images, charts, and complex layouts in long PDFs, often leading to static trade-offs between evidence coverage, noise, and inference cost. MAGE-RAG uses page retrieval as an entry point, building an offline evidence graph with page and element nodes encoding various relations like containment, reading order, and semantic neighbors. At query time, an online evidence controller iteratively activates, opens, searches, and prunes evidence under explicit budgets, rendering a compact, relevant evidence subgraph for the Large Vision-Language Model (LVLM). Experiments show MAGE-RAG achieves 52.75 overall accuracy on LongDocURL and 53.26 accuracy with 51.19 F1 on MMLongBench-Doc, demonstrating improved balance between dispersed evidence coverage and context-noise control.

Key takeaway

For Machine Learning Engineers developing multimodal RAG systems for long documents, MAGE-RAG offers a robust approach to overcome context limitations and noise. You should consider implementing an adaptive graph-based evidence construction strategy to dynamically balance evidence coverage with inference costs. This method allows your LVLMs to consume compact, relevant information, significantly improving accuracy on complex QA tasks involving diverse document elements like text, tables, and images.

Key insights

MAGE-RAG uses an adaptive evidence graph and query-time control for efficient multimodal RAG in long documents.

Principles

• Multigranular evidence graphs improve context relevance.
• Adaptive evidence construction balances coverage and noise.
• Page retrieval can serve as an effective entry point.

Method

MAGE-RAG builds an offline evidence graph (page/element nodes, various relations). An online controller then iteratively activates, opens, searches, and prunes evidence under budget constraints to form a subgraph.

In practice

• Implement graph-based RAG for complex PDF QA.
• Design adaptive evidence controllers for budget management.
• Integrate page-level visual retrieval as a RAG starting point.

Topics

• Multimodal RAG
• Long-Document QA
• Evidence Graphs
• Large Vision-Language Models
• Information Retrieval
• Adaptive Retrieval

Code references

• laonuo2004/MAGE-RAG

Best for: Research Scientist, AI Engineer, AI Scientist, Machine Learning Engineer, NLP Engineer

Related on AIssential

• When Iterative RAG Beats Ideal Evidence: A Diagnostic Study in Scientific Multi-hop Question Answering — Iterative RAG's staged retrieval and reasoning process outperforms static ideal evidence in complex multi-hop QA.
• MCERF: Advancing Multimodal LLM Evaluation of Engineering Documentation with Enhanced Retrieval — Multimodal retrieval and adaptive reasoning significantly enhance LLM performance on complex engineering documentation QA.
• AI-assisted Protocol Information Extraction For Improved Accuracy and Efficiency in Clinical Trial Workflows — AI-assisted RAG significantly boosts clinical trial protocol data extraction accuracy and efficiency over standalone LLMs.
• X-MADAM-RAG: Diagnosing and Handling Chinese-English Evidence Conflict in Retrieval-Augmented Generation — Multilingual RAG systems face significant challenges with contradictory evidence, requiring specialized diagnostic tools and pipelines like X-MADAM-RAG.
• TechGraphRAG: An Agentic Graph-Augmented RAG Framework for Technical Literature Reasoning — Agentic RAG frameworks can enhance technical reasoning by integrating multi-step pipelines, knowledge graphs, and external search for robust evidence grounding.
• Reducing Hallucinations in Complex Question Answering using Simple Graph-based Retrieval-Augmented Generation (long version) — Using a lightweight graph structure in RAG systems significantly reduces LLM hallucinations and improves factual correctness.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Artificial Intelligence.