AIBRIEF
The serverless database is replacing its primary vector index, the data structure powering AI style similarity search, with a new general search engine called v3.
turbopuffer, a serverless database built for fast vector search, is rewriting itself around a general search engine and demoting the vector index that defined it.
In a September 30 post, engineer Dan Harrison describes the existing architecture: documents keyed by ClusterId and LocalId, together the ANN address. SPFresh moves documents between clusters; attribute and full-text indexes point to the same address. Harrison names three costs: duplication for multi-vector records, write amplification as addresses shift, and document-scan blocks tied to clusters of about 100-200 documents. An earlier FTS redesign moved postings from ANN-cluster partitions to fixed 256-document blocks; the company says the FTS index shrank 10x and queries ran up to 20x faster. Those numbers are vendor-reported.
v3, the new primary-index rewrite, passed all CI in September but, as of the post, regressed significantly against production. Tuning has just started; there is no rollout date, no pricing change, and no customer deployment claim. Cursor and Notion were early customers whose workloads validated the original architecture; the post does not place either on v3.
The "RIP, vector database" headline is the vendor's obituary. The actual story is narrower: a specialized capability becoming one feature inside a general search engine, with the rewrite still proving itself.