Intelligent Multi-Agent RAG System for Flood Emergency Public Information

Abstract

The article presents the development and evaluation of an intelligent information system designed to support rapid decision-making and public communication during flood emergencies. The system is based on a Retrieval-Augmented Generation (RAG) approach and integrates official sources, domain knowledge and user interaction through an automated messaging interface. The goal of the research is to provide accurate, structured and per­sonalized recommendations for citizens by combining document retrieval with natural language generation. The architecture of the system incorporates modules for data acquisition, semantic indexing, query routing, and response generation, supported by a formal mathematical description of vector search and graph-based state transitions. A multi-agent design is applied to separate functional roles: operational guidance, legal information, and general inquiries. Each agent processes incoming requests within its domain, ensuring correct classification of user queries and structured output. The system retrieves relevant information dynamically, generates responses and formats them according to predefined templates, which contributes to reducing errors and ambiguity. A series of experi­ments was conducted on a representative set of user scenarios, including evacuation instructions, legal regulations and general awareness questions. Quantitative evaluation demonstrated high accuracy of responses and low er­ror rates, while response time remained suitable for real-time use. Qualitative examples confirmed consistency of output structure and relevance to user requests. The results show that the proposed system improves the clarity, speed, and reliability of emergency information delivery and may be applied to other risk management contexts requiring personalized, source-based recommendations, while also outlining prospects for its future integration with simulation-based evacuation models.