To overcome the fragmentation of information in aquatic nutrition and the non-factual responses generated by large language models， this study proposes a framework for an aquatic nutrition recommendation system that integrates semantic vectorization with knowledge graph reasoning. Using DeepSeek， entities and relationships were automatically extracted from more than 60 papers related to aquatic nutrition， leading to the construction of a specialized knowledge graph containing 4 936 nodes and 4 515 edges. This graph not only represents the basic associations between aquatic ingredients and nutritional components， but also establishes safety-oriented dietary contraindication chains for metabolic diseases such as gout and diabetes.At the service layer， the system adopts a dual-engine architecture by combining a vector database for broad semantic matching with a graph database capable of multi-hop reasoning. This design helps overcome the limitations of traditional single-point query methods. The experimental results show that the proposed hybrid information retrieval system can maintain the accuracy of recommendation results while enhancing the evidence-based medical support of the generated content. Empirical findings further indicate that the integration of semantic representation and knowledge graph reasoning improves the accuracy and credibility of nutrition plans for aquatic products. Overall， this study provides technical and data support for the development of aquatic nutrition regulation systems and offers a useful reference for related research.