Research on large model agent technology of geospatial information question answering based on water conservancy map

At present, generative artificial intelligence technology represented by natural language large model has ushered in vigorous development, providing a new way to obtain geospatial information. However, existing large models and agents equipped with large models, most of which are trained on general datasets, are highly prone to issues such as hallucinations, such as low correlation of answer content, inaccurate answer and lack of real-time performance. To address these issues, this paper proposes a technical framework for a large language model intelligent agent for geospatial information Q&A based on “one map” of water conservancy (GLMA). This technical framework integrates the agent's natural language understanding capabilities, enabling it to accurately grasp the user's question intent and subsequently drive the large model to execute specific "Water Resources Single Map" tasks. To enhance the accuracy of the task allocating, a tree-like structure task allocation system is used in the task dispatch stage, significantly improving the capabilities for task retrieval and parameter generation. Additionally, to validate the effectiveness of GLMA, this paper constructs a dataset for geospatial information question answering and establishes corresponding evaluation metrics. Compared with state-of-the-art Chinese open-source models such as Baichuan2, Llama3.1, ChatGLM4, and Qwen2.5, GLMA achieves the best results in terms of task allocation accuracy and query result accuracy. This research possesses significant extensibility and will lay the foundation for further studies on large model agents in other business domains.