Calculator
RAG Cost Estimator
If you are deciding whether to build retrieval yourself, the model bill is rarely the number that surprises people. Re-indexing is. This adds up embeddings, storage, refresh and inference so you can see which one dominates before you commit.
- Every component pricedNot just the model call
- Re-indexing includedThe cost people forget
- Runs in your browserNothing is sent anywhere
Work it out
Your numbers
Result
- Embeddings —
- Vector storage —
- Re-indexing —
- Inference —
- Cost per conversation —
- Largest component —
Rates assume roughly $0.02 per million embedding tokens, $0.25 per GB-month of vector storage, and $3 per million output tokens. Providers change these often — treat the shape of the answer as the useful part, not the decimal.
What the numbers mean
Re-indexing is the cost people miss
Embedding a corpus once is cheap. Embedding it again every time content changes is what turns a small bill into a recurring one. Set the refresh slider to what your content actually does, not what you hope it does.
Inference dominates once traffic is real
At low volume the index dominates and building it yourself looks cheap. At real traffic, generation takes over, and that is the point where the maths stops favouring a self-build.
This excludes the expensive part
None of this counts engineering time, evaluation, monitoring or the on-call rota. Those usually cost more than the infrastructure and they never appear in a comparison.
Cheaper is not the same as better
A self-built pipeline you cannot evaluate is worse than a managed one you can. Decide on the evaluation story first and let cost be the tiebreak.
Questions
They are representative mid-2020s rates and they move constantly. Use the tool to see which component dominates; check your provider's page for the exact figure.
Vectors are small. A thousand pages is well under a gigabyte, so storage almost never decides anything. It is included because leaving it out invites the question.
No. It means you should know which line grows with traffic before you decide. If the answer is inference, a managed product is competing on price, not just effort.
Want this run against your actual site?
Tell us the page your assistant lives on and the questions you care about. A person runs them and sends back the raw answers.