How Microsoft powers planet-scale AI apps with DiskANN
summary
DiskANN is a hybrid index that stores compressed vectors in RAM via product quantization and keeps uncompressed vectors and a graph on high‑speed SSDs inside Azure Cosmos DB, enabling cost‑effective planet‑scale vector search. Benchmarks show 10 million 768‑D vectors returning results in under 40 ms and P99 RU charges rising from ~30 to ~70 as scale grows; a 10M‑vector serverless ingest costs roughly $163, storage ~$23/month and 1M searches about $15. Upcoming features include sharded DiskANN and filter DiskANN for faster, tenant‑isolated and filtered searches.
AI-written summary · about
chapters (13)
- 0:00DiskANN and vector search
- 1:09Traditional vector indexing challenges
- 2:33DiskANN scalability and cost
- 3:36Azure Cosmos DB vector search role
- 5:42Cosmos DB fits AI apps
- 7:09Vector search in Cosmos DB apps
- 8:34Vector search queries in Cosmos DB
- 11:03Vector search performance metrics
- 12:31Vector search cost analysis
- 13:34Scaling to billions of vectors
- 15:20Upcoming optimizations and new features
- 17:09Hybrid search features
- 18:34Get started with vector search
related talks
11mScalable RAG with Azure Cosmos DB and DiskANN
48mThe power of AI and Copilot for Azure Databases
50mHow to design and build AI applications with vector search using Azure OpenAI & Azure Cosmos DB
40mUnleashing the Power of Vector Search with Azure SQL Database
03h 12mAzure Cosmos DB Conf 2024
12mBuilding Generative AI Apps with Azure Cosmos DB
27mKeynote
46mAdvanced RAG with Azure AI Search - top questions answered