With the exponential growth of video content on the internet, efficient retrieval of relevant videos has become a critical challenge. Traditional retrieval systems, which rely on keywords or metadata, often fail to address user intent and the semantic complexity of content, especially in big data environments. This paper presents the design and implementation of a novel video semantic retrieval system that integrates advanced video semantic understanding, similarity search using vector databases, and natural language processing powered by large language models. By embedding high-level semantic information and leveraging Approximate Nearest Neighbor (ANN) techniques, the system achieves accurate and efficient video retrieval. The proposed framework significantly enhances user experience by bridging the gap between user queries and video content semantics, offering new perspectives for applications in intelligent video search and recommendation systems. Experimental results demonstrate superior performance in accuracy and correlation compared to traditional retrieval methods, validating the effectiveness of the proposed approach.