VecFlow-Chamfer: A GPU-based Data Management System for High-Performance Multi-Vector Search on Superchips

Many emerging AI applications demand retrieval systems that go beyond document-level relevance and capture token-level semantics. Multi-vector search addresses this need through fine-grained semantic matching between token-level embeddings of queries and documents. However, it introduces significant system challenges, including compute-intensive set-to-set scoring, complex candidate filtering, and high memory overhead from storing dense token-level embeddings. Prior systems mitigate these challenges through GPU-based similarity calculation and indexing structures (e.g., IVFPQ-GPU) but often at the cost of reduced retrieval accuracy and suffer from low GPU utilization. We present \name, a GPU-based vector data management system that enables low-latency and high-recall multi-vector search on modern Superchip architectures. \name achieves this through a combination of three novel optimizations: (1) \maxivf, a GPU-native, compression-free index tailored for multi-vector search with fine-grained anchor vectors enabling scalable index construction and low-latency, high-accuracy candidate generation via CAGRA-based GPU routing; (2) \chamferkernel, a highly-optimized GPU kernel that enables single-digit millisecond Chamfer scoring over tens of thousands of candidate documents; and (3) \zerocomp, a tiered vector storage layer that supports on-demand, low-latency access to full-precision embeddings across Grace-Hopper NVLink-C2C interconnects. Together, these techniques enable \name to perform multi-vector search over hundreds of millions of document token embeddings with unprecedented high recall and low latency. Compared to state-of-the-art systems like PLAID and MUVERA, \name achieves an order-of-magnitude lower latency while significantly improving recall.