turbopuffer v3 drops ANN as primary index, hits 100% of CI passes

turbopuffer's original design, v1, stored only an ID and a vector per document, keyed by a combination of cluster ID and local ID called the ANN address. That layout worked well for vector search on object storage, and early customers including Cursor and Notion validated the tradeoffs. Over time, turbopuffer added attribute filtering and full-text search in v2, and customers like Linear began using it for non-search tasks such as a syncing engine. But because everything in a document was stored under its ANN address, the company says it faced three constraints: storage amplification, write amplification, and limited vectorization. The ANN cluster size of around 100–200 documents capped the block size for every query plan. By making ANN just another secondary index, turbopuffer hopes to remove that cap and speed up all query plans. The stakes hinge on whether the company can reach performance parity without regressing the vector search performance that its existing customers rely on.