LLM Agents Enhance Geospatial Data Retrieval with Risk Awareness

This research introduces an innovative framework that leverages large language models to streamline the retrieval of geospatial data from cloud-based catalogs. The system allows users to query complex remote sensing and environmental datasets using natural language, which is then translated into precise API calls. This significantly simplifies access to vast amounts of Earth observation data. The architecture is designed with three specialized agents: a Guardrail agent for enforcing safety and policy, a General-QA agent for interpreting user intent, and a Recommender-Analyst agent for generating schema-aware API calls. This modular and coordinated approach ensures robust and accurate interaction with external data services, making the framework portable across various platforms through API schema substitution. Preliminary adversarial evaluations indicate that while prompt-level safety instructions improve the system's resilience, rare but high-impact failures can still occur, particularly in API manipulation scenarios. This highlights the ongoing need for advanced, system-level defenses that can effectively balance safety, usability, and cost efficiency in such intelligent agent systems.