Multi-Agent RAG-Based Framework for Personalized Interaction with Large Language Models
Qian Wang, Wuxuan Duan, Haiyan Wang, Zhirong Luan, Yu Wang, Yulu Wang · 2025
With the breakthrough advancements of large language models (LLMs) in the field of natural language processing, their applications in tasks such as text generation and understanding have become increasingly widespread. However, in practical industrial management scenarios-particularly within traditional sectors such as the electric power industry-tasks such as information integration and data summarization still largely depend on manual processing or conventional human-computer interaction methods. These approaches commonly suffer from low summarization efficiency, heavy repetitive workload, high communication costs, and inaccurate information transmission. This paper proposes a personalized interaction method for large language models based on a multi-agent RAG framework. By incorporating hierarchical text agents, the framework enables more efficient and accurate integration of industry-specific information and the generation of personalized work summaries. It employs a hybrid mechanism of fixed and dynamic agents to generate text content tailored to the needs of multi-level managers in response to user feedback. An industry-specific dataset, ElePWL (Electric Power Work Logs), was constructed, and leading Chinese LLMs-ChatGLM3, DeepSeek-V3, and Tongyi Qianwen-Turbo-were deployed for ablation experiments. Control and experimental groups were established to evaluate the semantic similarity between generated summaries and reference texts. The results demonstrate that the proposed framework consistently improves the semantic alignment of personalized reports with standard references.