An AI-Driven Workflow Framework for Real-Time Disaster Response with Multi-Source Data Integration

Abstract

This study proposes an AI-driven workflow framework for integrating multiple data sources in disaster response. The framework was developed using the practical needs and challenges of the Taipei Water Department (TWD) as a real-world case. It addresses problems such as slow manual data collection, fragmented information, and unreliable AI-generated responses. The framework uses the PostgreSQL LISTEN/NOTIFY mechanism to receive and process fire alerts within milliseconds. This reduces system delay and database load. It also combines XPath-based web crawling with large language models (LLMs) to collect news from multiple sources. The framework checks and filters the information, then produces concise and reliable disaster summaries. To support accurate decision-making, the framework includes an automated OpenAI Vector Store update process. It organizes and preprocesses data, and uses vector-based retrieval to ensure responses are based on verified and up-to-date information. This helps reduce errors and outdated content. The framework also organizes data across different time periods, such as events within a 24-hour window. This supports short-, medium-, and long-term disaster monitoring. The deployed framework demonstrates the integration of official alerts with media reports in practical disaster response operations. By delivering maps and summaries through messaging bots, it improves collaboration and response efficiency at the disaster response centers.