pgvector on your own machine
Embeddings are big.Storage should not be the bill.
A modest retrieval setup is tens of gigabytes of vectors before the application has done anything interesting, and on a metered platform that is the line that grows every month. Lathe runs PostgreSQL 17 with pgvector on a dedicated machine whose disk is part of the price, so the size of your index is a capacity question, not a billing one.
Why vectors fit a machine with the storage included.
- pgvector from the first boot
- Enabled on every instance, with pg_stat_statements, pg_trgm and pgcrypto beside it. Create the extension, create the index, load the vectors.
- Storage sized with the plan, never metered
- The disk is NVMe on the machine itself. Fill it with vectors and the bill does not move; when it is nearly full you are told, and the next plan has more.
- Traffic for loading and re-embedding
- Bulk loads, re-embedding runs and exports move a lot of bytes. 20 TB a month is included, so a re-index is a job, not an invoice.
- Memory you can budget
- Vector search likes RAM. The plan's memory is yours alone; with Redis on the same machine, the budget between them is a slider you set.
- Backups that suit rebuildable data
- A daily image, 7 kept, restored in place. Embeddings can be regenerated from the source documents, which is exactly the kind of data a daily backup fits.
- Honest about the limit
- There is no point-in-time restore. For vectors that is rarely the constraint; for the transactional tables beside them, decide with that in mind.
Good for
- Retrieval for an AI assistant over a document set
- Semantic search over a product catalogue or a knowledge base
- An embeddings store beside the application's own tables, in one database
- Experiments that need tens of gigabytes of vectors without a storage meter
Questions
Which index types does pgvector support here?
The upstream pgvector build, so HNSW and IVFFlat, with the distance operators you would expect. The instance runs the unmodified extension.
How big can the index get?
As big as the plan's disk. The largest plan has the most; move up when you need more and the disk grows with it. Nothing is charged per gigabyte.
Is a dedicated machine overkill for a small vector store?
Mini starts at $15 a month with a 3-day trial and no card. It is the smallest machine we run, and it is still a machine of your own.
Can I keep a cache next to the vectors?
Yes. Redis runs on the same machine at the same price, with its own memory budget, for query results or session state.
Read next: Flat-price managed database, No egress fees, Never paused, never metered, Self-hosted VPS or managed, Supabase alternative, Neon alternative, Moving a Postgres database, Redis beside Postgres, A database for side projects, One machine, every engine. Every number on this page comes from the pricing page's facts.
pgvector on a machine of your own, from $15 a month.
Live about 90 seconds after checkout. Cancel any time; Mini starts with a 3-day free trial.
Create an instance