Dimensions are the biggest lever you control
Storage scales linearly with dimensions. Dropping from 1536 to 768 halves your bill, and many modern embedding models support dimension truncation (Matryoshka embeddings) with only a small retrieval-quality loss. Test recall at 512 and 768 before defaulting to the maximum.
Quantisation is nearly free money
int8 quantisation cuts storage by about 75% versus float32, with a typical recall drop of 1–3%. Binary quantisation goes further (32× smaller) and is often paired with a float rescoring pass over the top candidates — the standard trick for very large indexes.
- • float32 → int8: ~75% storage reduction, 1–3% recall loss.
- • Binary + rescore: up to 32× smaller, near-parity recall.
- • Fewer replicas in non-critical environments.
- • Prune stale chunks — indexes only ever grow if you let them.
Don't forget the index-build bill
Embedding 400,000 chunks at 400 tokens each is 160 million tokens — cheap on a small embedding model, but not free, and you pay it again every time you change chunking strategy or switch embedding models. Settle your chunking before you index at scale.
Related guides
Long-form playbooks on the same topic, written by the RevenueLab editorial team.
FAQ
How much does a vector database cost?
A typical startup index — 50,000 documents at 8 chunks each, 1536 dimensions, 2 replicas, 2M queries/month — runs roughly $20–$40/month on managed pricing, plus a one-off index build of a few dollars. Enterprise indexes with millions of documents reach hundreds to thousands per month.
How much storage does 1 million vectors need?
At 1536 dimensions in float32: 1M × 1536 × 4 bytes ≈ 5.7 GB raw, or roughly 8 GB with index and metadata overhead. Multiply by your replica count.
Is pgvector cheaper than a managed vector DB?
Usually yes at small scale, since you're already paying for Postgres. The crossover is typically somewhere in the low millions of vectors, where dedicated ANN indexes and horizontal scaling start to matter more than the licence saving.
How do I reduce vector database costs?
In order: reduce dimensions, quantise to int8, cut replicas in non-production environments, prune stale chunks, and cache repeated queries. Dimensions and quantisation together commonly cut storage 80%+.
Should I count query-time embedding costs?
Yes, but they're small — a 30-token query at $0.02/M costs $0.0000006. At 100M queries/month it becomes real money; below 10M it's a rounding error.
What chunk size should I use?
300–600 tokens with 10–20% overlap is the common default for prose. Smaller chunks improve precision but multiply vector count (and therefore storage) proportionally.
How this calculator is built
Independently maintained
Written by Sam Doshi and the RevenueLab editorial team. We don't sell the data feeds this tool is built on.
Sourced from primary data
Benchmarks come from public AdSense / Stripe / IRS disclosures and reader-submitted data — never third-party "$X per view" claims. Full methodology.
Last editorial review
Reviewed on a rolling quarterly cycle. Dated reviews are published on the methodology record for each calculator.
Editorial standards
See our editorial policy and disclaimer. Results are estimates, not advice.