What is DiskANN and when should you use disk-based vector search?

DiskANN is a graph-based ANN algorithm from Microsoft designed to serve billion-scale vector indexes from SSD instead of RAM, reducing memory cost by 10x or more. It uses a modified graph structure called Vamana that keeps good recall with lower connectivity than HNSW, and a compressed in-memory quantized copy for initial routing before fetching full vectors from disk. Use DiskANN when your dataset has billions of vectors and you cannot afford to keep them all in memory. Turbopuffer and Azure use DiskANN-style designs.