An Agentic AI Pipeline for an Evidence-Grounded Environmental Health Public Information System: Development and Evaluation

Background: Existing environmental and health information systems typically provide isolated data streams, requiring users to interpret heterogeneous information sources without personalized guidance. Residents and health providers of urban regions with complex environmental conditions, such as the Salt Lake Valley, frequently seek timely, location-specific information about weather, air quality, and their associated health impacts, but may find it difficult to integrate this information in their decision-making, precluding the effective translation of valuable environmental health information into action. Objective: This study aimed to develop and evaluate an agentic artificial intelligence (AI) pipeline that delivers evidence-grounded, neighborhood-scale responses to health-related environmental questions raised lay persons by dynamically orchestrating authoritative weather, neighborhood-scale environmental observations, air quality, and health-evidence resources. Methods: We developed an agentic AI pipeline using LangGraph and OpenAI large language models to interpret natural language queries, select and invoke relevant external tools, retrieve real-time environmental data and evidence-based health guidance, and synthesize responses. The pipeline integrates multiple authoritative resources, including weather forecasts, air-quality observations, geospatial services, and a retrieval-augmented generation (RAG) module built from asthma-related clinical guidance. The system incorporates fallback mechanisms to improve robustness when external services are unavailable. We evaluated the pipeline using a curated benchmark of representative user questions, assessing tool-needed recognition, correct tool selection, avoidance of irrelevant tools, parameter mapping, analytic pathway selection, and incorporation of tool outputs into the final response. Results: The developed system successfully integrated heterogeneous environmental and health information sources into a unified conversational interface that produces evidence-grounded, neighborhood-scale responses for the Salt Lake Valley. Evaluation demonstrated generally strong tool recognition, parameter mapping, and incorporation of retrieved information across diverse question types, supporting the pipeline's ability to retrieve and integrate authoritative environmental and health information consumable by lay persons and providing an important step for the T3 and T4 phases of the translational pipeline of exposure health informatics. Conclusions: By combining existing environmental health knowledge with local sensor observations and neighborhood-scale environmental measurements, and authoritative state and national weather and air quality resources, the system delivers more localized and context-aware guidance than approaches that rely on authoritative regional or state-level information. Rather than depending on the internal knowledge of a large language model, the pipeline dynamically retrieves, synthesizes, and attributes evidence from authoritative sources. The modular architecture enables seamless integration of additional environmental data streams, health knowledge sources, and geographic regions, providing a scalable foundation for future translational pipelines, completing the T3 to T4 path for environmental health.