Understand the bill
What RAG actually costs (and what it doesn't)
Retrieval-augmented generation has a reputation for being expensive — mostly earned by bad estimates. The pipeline has four cost lines: one-time indexing (embedding your corpus), vector database (storing vectors and searching them), recurring embeddings (each query, plus re-indexing new content), and LLM generation (writing the answers). The calculator prices each line separately, because they behave completely differently at scale.
The counterintuitive part: indexing is cheap
Embedding a 5,000-page corpus with a modern embedding model costs under a dollar — the defaults above work out to about 6 cents. Storage is similarly small: those pages become ~3,600 vectors occupying a few hundredths of a GB. The myth that "indexing our documents will be expensive" rarely survives contact with the arithmetic. What actually costs money is answering questions: LLM generation is typically 90%+ of a RAG bill at production query volumes.
Where the money goes at scale
- Retrieved context is input tokens. Every chunk you paste into the prompt is billed on every query. Retrieving 5 chunks when 3 suffice inflates the dominant cost line by ~40%.
- Answer length is the second lever. Grounded answers with citations don't need 800 tokens; cap them.
- Vector DB pricing units vary wildly. Pinecone bills read/write units, others bill pods or compute hours. That's why this tool exposes the rates as editable fields — put in your provider's effective numbers, starting from serverless list prices like Pinecone's.
Embedding model choice
Small embedding models ($0.02/1M tokens, 1,536 dimensions) handle most retrieval tasks; large ones ($0.13/1M, 3,072 dims) buy marginally better recall at 6× the embedding cost and double the storage, because storage scales with dimensions. Start small; upgrade only if retrieval quality measurably fails your evals.
Worked example
The defaults: 5,000 pages indexed for ~$0.06 one-time, stored in ~0.03 GB. At 50,000 queries/month with GPT-5.4 mini generating answers: vector DB ~$1.20, embeddings ~$0.03, generation ~$142 — ~$144/month total, 0.3¢ per query. The lesson is visible immediately: the database is noise, the LLM is the bill. If you're wrapping this pipeline in a chatbot, price the full product in the chatbot cost calculator, and check whether it pays off in the ROI calculator.