Multimodal RAG Survey

MLLMs with and without Multimodal RAG, and growth in related publications

Why Multimodal RAG

From context overflow to grounded document intelligence

Large documents — 1000+ pages of interleaved text, tables, and charts — overwhelm a model's context window. Multimodal RAG retrieves only the evidence that matters, enabling reasoning that native MLLMs cannot reach alone.

(a) MLLMs with vs. without Multimodal RAG for large-document comprehension. (b) Explosive growth in related publications from 2024 to 2025.

Papers in 2025

Growth vs. 2024

Abstract

What This Survey Covers

Document understanding is critical for applications from financial analysis to scientific discovery. Current approaches, whether OCR-based pipelines feeding Large Language Models (LLMs) or native Multimodal LLMs (MLLMs), face key limitations: the former loses structural detail, while the latter struggles with context modeling. Retrieval-Augmented Generation (RAG) helps ground models in external data, but documents' multimodal nature — combining text, tables, charts, and layout — demands a more advanced paradigm: Multimodal RAG. This approach enables holistic retrieval and reasoning across all modalities, unlocking comprehensive document intelligence. Recognizing its importance, this paper presents a systematic survey of Multimodal RAG for document understanding. We propose a taxonomy based on domain, retrieval modality, and granularity, and review advances involving graph structures and agentic frameworks. We also summarize key datasets, benchmarks, applications and industry deployment, and highlight open challenges in efficiency, fine-grained representation, and robustness, providing a roadmap for future progress in document AI.

Highlights

Key Contributions

Comprehensive Taxonomy

The first survey that explicitly bridges multimodal RAG and document understanding, with a taxonomy spanning domain (open/closed), retrieval modality (image/text/hybrid), granularity (page/element), and hybrid enhancements (graph- and agent-based).

Datasets, Benchmarks & Evaluation

A broad collection of multimodal RAG datasets, benchmarks, and comparative results for systematic evaluation, covering metrics for both retrieval-oriented and generation-oriented assessment.

Applications & Industry Deployment

Practical applications spanning finance, scientific literature, and social analysis, alongside industrial deployment considerations and real-world usage patterns.

Challenges & Future Directions

Open challenges in efficiency, fine-grained multimodal representation, and robustness — a concrete roadmap for future progress in document AI.

The Map

Taxonomy Overview

One color system runs through the whole page — learn it once, then scan it across the figures and tables below.

Open vs. Closed Domain

Open-domain RAG retrieves from large corpora; closed-domain RAG focuses on a single document, selecting only the most relevant pages to cut context length and hallucination.

Retrieval Modality

Image-based retrieval encodes pages as images via VLMs; image+text retrieval combines visual features with OCR or MLLM-generated captions for richer cross-modal representations.

Retrieval Granularity

From page-level retrieval (whole pages as units) to element-level retrieval (tables, charts, images, text blocks) for finer-grained evidence localization and grounding.

Graph-based RAG

Multimodal content as graphs where nodes are content units and edges encode semantic, spatial, and contextual relations for structured retrieval and reasoning.

Agent-based RAG

Autonomous agents formulate queries, select retrieval strategies, and adaptively fuse information across modalities with iterative reasoning and verification.

Hybrid Enhancements

Combining graph-based indexing with agent-based orchestration for more flexible, interpretable, and robust multimodal RAG systems.

Reference

BibTeX

@article{gao2025scaling,
  title={Scaling Beyond Context: A Survey of Multimodal Retrieval-Augmented Generation for Document Understanding},
  author={Sensen Gao and Shanshan Zhao and Xu Jiang and Lunhao Duan and Yong Xien Chng and Qing-Guo Chen and Weihua Luo and Kaifu Zhang and Jia-Wang Bian and Mingming Gong},
  year={2025},
  eprint={2510.15253},
  archivePrefix={arXiv},
  primaryClass={cs.CL},
  url={https://arxiv.org/abs/2510.15253},
}

Scaling Beyond Context: A Survey of Multimodal Retrieval-Augmented Generation for Document Understanding