A hierarchical multi-agent system for document question answering

By using a hierarchical multi-agent system, the problems of insufficient document parsing depth and limited ability to handle complex problems are solved, enabling the document question-answering system to be adaptive and generate high-quality answers, ensuring the continuous evolution of the system and meeting the personalized needs of users.

CN122412466APending Publication Date: 2026-07-17SHANGHAI UNIV
View PDF 1 Cites 0 Cited by

Patent Information

Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
SHANGHAI UNIV
Filing Date
2026-03-23
Publication Date
2026-07-17

AI Technical Summary

Technical Problem

Existing technologies for extracting information from unstructured or semi-structured documents and answering user questions suffer from insufficient document parsing depth, limited ability to handle complex questions, lack of verification loops in answer generation, and static system architecture, making it difficult to adapt to personalized user needs.

Method used

A hierarchical multi-agent system is adopted, including a perception and parsing layer, a reasoning and retrieval layer, a decision generation layer, and a verification and optimization layer. Through multi-format adaptation, content extraction, and semantic-level structuring, the system dynamically decomposes user questions, generates and optimizes answers, and performs multi-dimensional quality assessment to achieve system self-adaptation and continuous evolution.

Benefits of technology

It achieves a deep understanding of documents and accurate answers to various types of questions, ensuring high-quality and reliable answers. The system can adapt to user needs and continuously optimize.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN122412466A_ABST
    Figure CN122412466A_ABST
Patent Text Reader

Abstract

The application relates to a hierarchical multi-agent system for document question answering, comprising: a perception analysis layer: multi-format adaptation, content extraction and semantic structuring are carried out on original documents, standardized JSON and semantic packages are output, and the standardized JSON and the semantic packages are stored in a knowledge base; an inference retrieval layer: user question types are identified, including fact type, inference type, comparison type, multi-hop type and scheme type, a question is dynamically disassembled, a decision retrieval path is decided, knowledge segments are retrieved and sorted from the knowledge base and a history base, and Top-N results with trace information are output; a decision generation layer: an answer framework is constructed based on the Top-N knowledge segments, and an initial answer with source annotation is generated; a verification optimization layer: multi-dimensional quality evaluation and optimization are carried out on the initial answer, and a final answer is output; a sedimentation iteration layer: data of each layer is arranged, the knowledge base is updated based on full-amount data and the final answer, and an intelligent agent is optimized. Compared with the prior art, the application solves the technical bottleneck problems of a traditional document question answering system in complex document analysis, deep semantic inference and answer reliability guarantee and the like.
Need to check novelty before this filing date? Find Prior Art