RIP, vector database
- Source
- turbopuffer.com
- Author
- razin
- Date

Context
turbopuffer's planned v3 storage redesign separates a search engine's document storage from its vector index, which groups similar numerical representations for retrieval. In the existing layout, documents are stored under their vector's index address. Rebalancing those groups can therefore move document contents and references used by other search indexes too.
Making vector search a secondary index is intended to remove that coupling and give other operations, including aggregations, more freedom in how they read data. Earlier in September the new implementation passed all of the company's continuous-integration tests, but when the article appeared on September 30, performance tuning and production rollout were still ahead. The article establishes the design rationale rather than a completed speed improvement.
Checking sign-in…
Loading comments…