RAG cost estimator: indexing, queries and storage
Enter your document volumes and expected traffic to see the three cost lines of a retrieval-augmented knowledge base. Every assumption is shown below and editable.
- Runs in your browser, nothing is sent
- Illustrative prices you can overwrite
- Maths shown in full
Describe your corpus and traffic
Defaults describe a mid-sized internal knowledge base. Change any value; results recalculate as you type.
Tip: most teams route easy questions to an economy model and escalate the rest. Model that by running the tool twice with split traffic.
- One-off indexing (embedding)$0
- Monthly LLM query cost$0
- Monthly vector storage$0
- Monthly re-embedding (content churn)$0
- Cost per 1,000 answers$0
Size of the index
- Pages0
- Tokens in corpus0
- Vectors (chunks)0
- Approximate index size0 GB
- Tokens sent per answer (input + output)0
All figures are estimates built from the assumptions below.
Unit prices and parameters
Prices below are round placeholder figures in US dollars, not quotes from any vendor. Replace them with your actual price list. Dimensions decide storage; token prices decide monthly cost.
| Parameter | Value | Why it matters |
|---|---|---|
| Dense legal text runs higher; slide decks lower. | ||
| Overlap repeats text at chunk boundaries, raising vector count. | ||
| Multiplies input tokens per answer. Reranking lets you lower k safely. | ||
| Sent on every call. Prompt caching, where available, cuts this. | ||
| Also the tokens embedded per query. | ||
| Changed content is re-embedded; drives the monthly re-embedding line. | ||
| Illustrative. | ||
| Illustrative. | ||
| Illustrative. | ||
| Storage per vector = dimensions × 4 bytes (float32). | ||
| Illustrative. | ||
| Illustrative. | ||
| Illustrative. | ||
| Illustrative. | ||
| Illustrative. | ||
| Illustrative. | ||
| Illustrative. | ||
| Illustrative. | ||
| HNSW graphs and metadata add to raw vector size. | ||
| Illustrative. Self-hosted pgvector may be near zero marginal cost; managed services vary widely. | ||
| Use working days if traffic is weekday-only. |
How the maths works
Nothing clever happens behind the panel. The formulas are the ones we use in scoping calls, written out so you can challenge them.
pages = documents × pages per document
tokens = pages × tokens per page
stride = chunk size × (1 − overlap)
vectors = ceil(tokens ÷ stride)
indexing = vectors × chunk size ÷ 1,000,000 × embedding price
index size = vectors × dimensions × 4 bytes × overhead factor
storage/mo = index size (GB) × storage price
input/answer = system prompt + question + k × chunk size
output/answer= answer length
cost/answer = input ÷ 1M × input price + output ÷ 1M × output price
+ question ÷ 1M × embedding price
queries/mo = cost/answer × questions per day × days
re-embed/mo = indexing × content changed per month
monthly = queries/mo + storage/mo + re-embed/moWhat is not included
- Parsing and OCR of scanned documents, which can exceed embedding cost for image-heavy archives.
- A reranker model, typically priced per query or per 1,000 documents scored.
- Application hosting, logging and evaluation tooling; see LLMOps and monitoring.
- People: content owners, reviewers and the team that keeps the test set honest.
Reading the result
For most corpora the one-off indexing cost is small and the monthly query cost dominates. That is why chunk size, k and model tier are the levers worth arguing about, and why reranking often pays for itself by letting you send fewer, better chunks. Storage only becomes significant past tens of millions of vectors, which is also the point where database choice starts to matter.
Turn the estimate into a plan
Enterprise RAG
Reference architecture, connectors and evals for a production knowledge base.
ExploreVector database chooser
Eight questions, a ranked recommendation with reasons.
Try itLLM cost calculator
Model monthly spend for any LLM application, not only RAG.
OpenRAG explained
The mechanism behind every line in this estimate.
ReadFrequently asked questions
Where do the default unit prices come from?
They are illustrative round numbers chosen to sit in the range of common API and managed-database pricing at the time of writing. They are not quotes from any provider. Edit the assumptions table to match the price list you actually use; the results update immediately.
What costs does the estimator leave out?
Document parsing and OCR, a reranker model if you use one, application hosting, observability tooling and people. Parsing scanned PDFs in particular can cost more than embedding them. Our guide to the cost of LLM applications covers the full picture.
Why does chunk size change the monthly cost so much?
Every answer sends k retrieved chunks to the model, so input tokens per query scale with chunk size times k. Halving chunk size with the same k roughly halves the LLM input cost, but smaller chunks can lose context. Tune both against a labelled test set; see embeddings strategy.
Is the storage figure the bill from the vector database?
It is an approximation: vectors times dimensions times four bytes, multiplied by an index overhead factor, priced per GB-month. Managed services price in different units (read units, pods, storage tiers), so treat it as an order of magnitude and confirm with the provider. The selection guide compares cost models qualitatively.
Can you help us reduce these numbers?
Yes. The usual levers are smaller or cached system prompts, fewer and better-ranked chunks through reranking, a cheaper model for easy questions with routing to a stronger one, and quantised vectors. Send us your estimate and we will review it with you.