Introducing AISAQ in Milvus: Billion-Scale Vector Search Just Got 3,200× Cheaper on Memory

What this post added

Introduces AISAQ, a new disk-based vector index for Milvus that achieves a 'Zero-DRAM-Footprint Architecture' by storing PQ codes and other search-critical data on disk. This significantly reduces memory usage for billion-scale vector search workloads, offering a 3,200x reduction in memory footprint compared to previous approaches while maintaining performance. The post explains the underlying principles of graph-based vector search, the trade-offs involved, and how AISAQ builds upon DISKANN by optimizing disk data layout for efficient I/O.