Plain Large Language Models (LLMs) often struggle with hallucinations and complex multi-objective reasoning when applied to Geographic Artificial Intelligence (GeoAI). To address these limitations, we present GeoAgent-MG-RAG, which is a multi-agent framework that integrates a dynamic Geospatial Multi-modal Knowledge Graph (GeoMMKG) with Retrieval-Augmented Generation (RAG). We have evaluated our framework on a heterogeneous and comprehensive multi-modal dataset of Chicago and NYC in USA that is consisting of thousands of graph nodes and Points of Interest (POIs), and raster data files (e.g., NDVI, DEM), in addition to other heterogeneous sources (e.g., OSM, AlphaEarth and hyperlocal air quality data), and GeoAgent-MG-RAG exhibited a robust performance in orchestrating complex urban workflows. Also, we have performed a comprehensive testing on four foundation models (i.e., Llama 4, GPT OSS, Kimi K2, Qwen3) using 160 natural language queries, and our framework achieves a peak tool-calling accuracy of Approximately 97.5% using few-shot learning, thus demonstrating a noticeable improvement over standalone baselines. Our system also successfully identifies Pareto-optimal solutions and balanced the distance, pollution exposure, and scenic value for the task of multi-objective route planning. Also, for the POI itineraries recommendation task, our system maintains high positional relevance (i.e., NDCG (Formula presented) ) even as complexity increases. These findings position GeoAgent-MG-RAG as a scalable approach for mitigating the inherent plain LLM limitations in smart city analytics, thus closing the gap between geospatial cognitive reasoning and actionable, evidence-based urban intelligence.
Al Jawarneh, I.M., Bagosher, M.O., Foschini, L., Bellavista, P. (2026). GeoAgent-MG-RAG: A multi-agent LLM framework with spatial knowledge graphs for context-aware urban planning and health analytics. INFORMATION PROCESSING & MANAGEMENT, 64(1), 1-40 [10.1016/j.ipm.2026.105086].
GeoAgent-MG-RAG: A multi-agent LLM framework with spatial knowledge graphs for context-aware urban planning and health analytics
Foschini, Luca;Bellavista, Paolo
2026
Abstract
Plain Large Language Models (LLMs) often struggle with hallucinations and complex multi-objective reasoning when applied to Geographic Artificial Intelligence (GeoAI). To address these limitations, we present GeoAgent-MG-RAG, which is a multi-agent framework that integrates a dynamic Geospatial Multi-modal Knowledge Graph (GeoMMKG) with Retrieval-Augmented Generation (RAG). We have evaluated our framework on a heterogeneous and comprehensive multi-modal dataset of Chicago and NYC in USA that is consisting of thousands of graph nodes and Points of Interest (POIs), and raster data files (e.g., NDVI, DEM), in addition to other heterogeneous sources (e.g., OSM, AlphaEarth and hyperlocal air quality data), and GeoAgent-MG-RAG exhibited a robust performance in orchestrating complex urban workflows. Also, we have performed a comprehensive testing on four foundation models (i.e., Llama 4, GPT OSS, Kimi K2, Qwen3) using 160 natural language queries, and our framework achieves a peak tool-calling accuracy of Approximately 97.5% using few-shot learning, thus demonstrating a noticeable improvement over standalone baselines. Our system also successfully identifies Pareto-optimal solutions and balanced the distance, pollution exposure, and scenic value for the task of multi-objective route planning. Also, for the POI itineraries recommendation task, our system maintains high positional relevance (i.e., NDCG (Formula presented) ) even as complexity increases. These findings position GeoAgent-MG-RAG as a scalable approach for mitigating the inherent plain LLM limitations in smart city analytics, thus closing the gap between geospatial cognitive reasoning and actionable, evidence-based urban intelligence.I documenti in IRIS sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.



