Two years of vector search at Notion: 10x scale, 1/10th cost
Summary
Notion shares how it scaled vector search for Notion AI from launch to millions of workspaces, achieving a 10x scale and 90% cost reduction by migrating from dedicated pods to serverless embeddings and adopting turbopuffer with Ray/Anyscale. The post covers architecture, indexing strategies, cost savings, and ongoing optimization to reduce latency and data volume, with future plans like expanding data sources and Notion Agents.