Abstract: Against the backdrop of the ongoing development of the national water network and advances in artificial intelligence, there is an urgent need for large-scale hydraulic dispatching models to support digital empowerment. This need arises from the complexity of water diversion systems, which feature numerous sluice-pump facilities, densely distributed hydraulic dispatching nodes, strong coupling among real-time dispatching operations, massive data volumes, and diverse operational scenarios. To address technical challenges such as multi-source data integration, domain knowledge adaptation, and dynamic decision-making support, this study develops a framework for a large-scale hydraulic dispatching model for the Eastern and Middle Routes of the South-to-North Water Diversion Project (SNWDP). Based on multimodal coupling and retrieval-augmented generation (RAG), the framework follows a structure of data foundation, model development, and scenario adaptation. It comprises three key technologies: an L0 multimodal data foundation for deep data fusion and efficient access; an L1-L2 model-RAG integration mechanism driven jointly by data and models; and an L3 panoramic digital twin platform for multi-scenario adaptation. Together, these technologies enable multimodal data fusion, scenario adaptation, and scheduling scheme generation. The framework was applied to three representative scenarios: hydraulic dispatching for the main canal of the Middle Route of the SNWDP under along-channel water-level constraints; hydraulic dispatching for the main canal of the Middle Route considering water-level fluctuations of conveyance structures; and water grid hydraulic dispatching for the Eastern Route considering flow demand and water quality constraints. The results indicate synergistic effects among the framework modules. Compared with conventional L1 and L2 models, the RAG-enabled hydraulic dispatching model improves interpretability and real-time responsiveness. Its comprehensive dispatching accuracy for hydrodynamics and water quality reaches 0.845, which is 0.098 higher than that of conventional open-source RAG systems. This framework is applicable to large-scale water diversion projects and provides a feasible technical pathway for their high-quality digital and intelligent development.