<p>Inspection of hydraulic projects is a core management task to ensure the safe and stable operation of major infrastructure. The current traditional inspection model has significant shortcomings, such as high reliance on manual labor, low accuracy in identifying potential risks, and insufficient dynamic decision-making capabilities. This paper innovatively constructs a multi-agent collaborative intelligent decision-making framework that integrates large language models (LLMs) and graph retrieval-augmented generation (Graph RAG) technologies. Through a modular architecture encompassing perception, cognition, and decision, the framework achieves full-process automated inspection. The study employs multi-source inspection data from the past three years to construct a multimodal dataset. Domain-adaptive fine-tuning enhances the F1 scores of the multimodal large model in equipment recognition and defect detection by 7.2% and 6.9%, respectively. Furthermore, a dynamic knowledge graph system based on Graph RAG is established. Knowledge injection techniques compensate for gaps in domain-specific knowledge, while entity-relation reasoning mechanisms effectively mitigate model hallucination phenomena. Experimental results demonstrate that the hydraulic engineering inspection reports generated by this method, evaluated by both experts and operations personnel, accurately reflect professional knowledge and technical depth in the field of hydraulic engineering maintenance. This research provides a new technical paradigm with strong explainability and high reliability for the intelligent operation and maintenance of hydraulic engineering infrastructure, offering significant engineering application value to promote digital transformation within the industry.</p>