GeoAgentic-RAG: A Multi-Agent framework for autonomous geospatial reasoning and visual insight generation with LLM

• GeoAgentic-RAG unites multi-agent systems and RAG for geospatial intelligence. • Autonomous spatial reasoning and visual insight generation enabled by LLMs. • Unified framework supporting vector, raster, and multimodal geospatial data. • Experiments demonstrate strong semantic consistency with 0.76 relevance score. • Enhanced efficiency and accuracy in geospatial information retrieval and analysis. Conventional Retrieval-Augmented Generation (RAG) systems have limited effectiveness in geospatial question answering because text-based similarity retrieval cannot adequately represent spatial semantics such as topology and spatial context. To overcome this limitation, we propose GeoAgentic-RAG, a multi-agent framework that enables multimodal large language models (MLLMs) to perform autonomous geospatial reasoning. The framework integrates natural language query parsing, semantic-spatial retrieval, and executable geospatial analysis within a unified, agent-based workflow. Multiple specialized agents collaboratively interpret user queries, retrieve relevant vector and raster datasets from a unified geospatial database, decompose analytical tasks, generate valid spatial logic, and produce interpretable analytical results. We evaluate GeoAgentic-RAG using a benchmark of geospatial retrieval, feature characterization, and spatial relational reasoning tasks in Nanjing and Guangzhou. The proposed framework achieves a pass rate of 85.3% and an answer correctness of 88.3%, outperforming conventional RAG methods and representative code-generation baselines. These results demonstrate that agent-based integration of retrieval and spatial analysis substantially improves the reliability of geospatial question answering and provides a practical framework for the next-generation intelligent GIS applications.