A Knowledge-Tool Collaborative Agent Construction Method for Public-Oriented Earthquake Emergency Decision Services

Objectives: Post-earthquake protective action and evacuation decisions for the general public are characterized by high urgency, short decision windows, and strong dependence on geospatial context. Existing science-popularization-oriented knowledge services often fail to provide actionable support tailored to an individual’s location and environmental constraints, while dialogue generation based solely on large language models (LLMs) lacks spatial computing capability and may suffer from hallucinations. To address this problem, this study proposes a public-oriented earthquake emergency information service agent framework and explores a controlled, tool-grounded pathway from natural-language needs to spatialized action recommendations. Methods: First, scenario templates and atomic emergency response measures were constructed from authoritative guidelines and emergency plan corpora. A scenario–measure semantic mapping was then established through vector-based retrieval to provide traceable knowledge constraints for recommendation generation. Second, targeting typical public geospatial analysis needs, GIS operations such as risk identification, facility retrieval, and evacuation route planning were abstracted into callable tools. A unified tool descriptor schema and registry mechanism were further designed to form an extensible tool space. On this basis, an end-to-end workflow was developed, including intent parsing, scenario matching, measure recommendation, toolchain planning, sequential execution, and response generation, thereby enabling coordinated knowledge retrieval, tool invocation, and result organization. A prototype system was also implemented to support the presentation of intermediate outputs and experiment reproduction. Results: The experimental evaluation included both a representative case demonstration and controlled multi-task assessment. The former was used to illustrate the full end-to-end process from user request interpretation to spatial result generation, while the latter involved 12 test samples covering four representative task types, namely risk identification, shelter lookup, full-chain evacuation, and family rendezvous. The controlled evaluation results indicate that, under the current experimental setting, the system performs relatively stably in intent recognition, task-template matching, and key tool execution, with the best overall performance observed in the full-chain evacuation task. At the same time, some tasks still reveal insufficient feedback organization from execution results to final response generation. Conclusions: Overall, this study preliminarily demonstrates the feasibility of a knowledge-constrained and tool-grounded collaborative route for public earthquake emergency decision support. The proposed framework provides an implementable pathway for transforming natural-language requests into spatial analysis results and response outputs under controlled conditions. Future work will focus on integrating real operational data, developing result-aware response organization mechanisms, and establishing reliability evaluation and user testing schemes.