Scholarlink

This study presents MMA-RAG, a multi-modal, multi-agent Retrieval-Augmented Generation (RAG) system tailored for insurance document processing. Our approach addresses challenges posed by diverse, unstructured data such as scanned PDFs, tables, and handwritten documents, which require structured retrieval for accurate AI-driven responses. By leveraging multi-modal models and agent-based decision-making, we integrate structured (Excel, SQL databases) and unstructured (OCR-extracted text, policy documents) data into a unified processing pipeline. The system achieves a high coherence rate of 98.56%, with improvements in retrieval precision and regulatory compliance. Human and automated evaluations demonstrate the effectiveness of the system in refining response quality. Additionally, we introduce specialized agents, namely a User Proxy Agent and a Reviewer Agent, to dynamically manage queries and verify output accuracy. Comparative analysis highlights the benefits of the proposed architecture compared to baseline RAG approaches. Finally, the study acknowledges the system’s limitations and proposes directions for future research.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

MMA-RAG: Multi-Modal Agents for Insurance Document Processing with Retrieval-Augmented Generation

  • 会议论文

  • 发表日期：

    2025年

查看全文
  • Ibrahim Krayem,
  • Malak Ghourabi,
  • Mohamad Al Assaad

摘要

This study presents MMA-RAG, a multi-modal, multi-agent Retrieval-Augmented Generation (RAG) system tailored for insurance document processing. Our approach addresses challenges posed by diverse, unstructured data such as scanned PDFs, tables, and handwritten documents, which require structured retrieval for accurate AI-driven responses. By leveraging multi-modal models and agent-based decision-making, we integrate structured (Excel, SQL databases) and unstructured (OCR-extracted text, policy documents) data into a unified processing pipeline. The system achieves a high coherence rate of 98.56%, with improvements in retrieval precision and regulatory compliance. Human and automated evaluations demonstrate the effectiveness of the system in refining response quality. Additionally, we introduce specialized agents, namely a User Proxy Agent and a Reviewer Agent, to dynamically manage queries and verify output accuracy. Comparative analysis highlights the benefits of the proposed architecture compared to baseline RAG approaches. Finally, the study acknowledges the system’s limitations and proposes directions for future research.