Abstract
Vector search plays a crucial role in many real-world applications. Inaddition to single-vector search, multi-vector search becomes important formulti-modal and multi-feature scenarios today. In a multi-vector database, eachrow is an item, each column represents a feature of items, and each cell is ahigh-dimensional vector. In multi-vector databases, the choice of indexes canhave a significant impact on performance. Although index tuning for relationaldatabases has been extensively studied, index tuning for multi-vector searchremains unclear and challenging. In this paper, we define multi-vector searchindex tuning and propose a framework to solve it. Specifically, given amulti-vector search workload, we develop algorithms to find indexes thatminimize latency and meet storage and recall constraints. Compared to thebaseline, our latency achieves 2.1X to 8.3X speedup.