With the rapid advancement of artificial intelligence and big data technologies, high-dimensional vector retrieval has become a fundamental technology in areas such as recommender systems and information retrieval. When facing billion-scale datasets, traditional graph-based approximate nearest neighbor search (ANNS) methods struggle to balance index construction efficiency, query performance, and system scalability due to single hardware limitations. This paper proposes a CPU-GPU collaborative pipelined vector retrieval framework targeting large-scale datasets. The framework addresses GPU memory constraints through collaborative clustering with dynamic load balancing and asynchronous double buffering. During graph construction, we employ a two-phase pipelined approach where GPU performs parallel k-NN computation while CPU simultaneously optimizes graph structures. For query processing, the system implements collaborative graph navigation with intelligent prefetching and adaptive workload distribution. Experimental results on billion-scale datasets demonstrate significant improvements: our method achieves sub-400 ms query latency with over 90,000 QPS, representing 150–200 times improvement over traditional approaches and 2–3 times better than BANG. On deep learning features, query latency is reduced by 50–70% while maintaining high recall accuracy.