Oral 1: The RaBitQ Library

High-dimensional vectors serve as a bridge between massive real-world data and AI applications such as large language models (LLMs). Yet, the data is often of large scale and with high dimensionality, which causes high costs in both memory consumption and computation. To reduce both costs, vector quantization has been widely adopted. A recent advance in vector quantization is the RaBitQ method, which supports flexible compression rates and provides asymptotically optimal error bounds for distance estimation. It demonstrates strong empirical performance in terms of space, accuracy and runtime. In this paper, we first provide a comprehensive overview of RaBitQ’s design, insights and theoretical guarantees. We then introduce the RaBitQ Library, which covers some key implementation techniques of RaBitQ and its integration with indices for approximate nearest neighbor search.