
A serverless vector and full-text search engine built on object storage, priced on usage with a monthly minimum — the cost-at-scale option behind many AI products' retrieval.
turbopuffer stores everything in object storage such as S3 and serves frequently accessed data from a memory and SSD cache, which is why its pitch is cost at very large scale rather than the lowest latency on a small index. It does vector, full-text, and hybrid search with metadata filtering through one API, and it is designed for very high namespace counts — one index per customer or per user. The homepage lists customers including Anthropic, Notion, Linear, Ramp, and Harvey, and claims more than a trillion documents in production; both are vendor-reported. Pricing is usage-based with a monthly minimum: Launch $16, Scale $256, and Enterprise $4,096 or more with a 35 percent usage premium, with storage at up to $0.33 per GB, writes up to $2 per GB, and queries billed on data queried plus data returned. A SOC 2 report is available on every plan, a HIPAA BAA on Scale and up, and BYOC on Enterprise only. There is no free tier. Because data that has fallen out of cache is read from object storage, benchmark first-query latency on idle namespaces before committing. Buy turbopuffer if you run many tenants or billions of documents and storage cost dominates the bill. Buy Pinecone for a mature managed service with a free starter tier, or Milvus if you need to self-host.
Get a new AI workflow each week — many feature turbopuffer and other tools in this category.