Staged Vector Stream Similarity Search Methods

This article describes the staged vector stream similarity search methods, or briefly SVS, designed to index and search vector streams by similarity over a time interval. SVS continuously adapts to the vector stream as the vectors are received and do not depend on costly updates on an index structure. The article presents experiments to investigate the performance of two implementations of SVS, one based on product quantization and another based on Hierarchical Navigable Small World graphs. Finally, the article describes a proof-of-concept implementation of a classified ad retrieval tool that uses staged HNSW on real data collected from an online classified ads company.