What does a production RAG app actually cost to run? Estimate the one-time ingest, the per-query LLM spend, and your monthly bill. All prices are editable — defaults are representative list prices, not gospel.
1. Ingest (one-time): you embed your whole corpus once. Cheap per token, but scales with corpus size. 2. Per-query embeddings: each question is embedded — tiny. 3. Per-query LLM: the retrieved chunks become input tokens, plus the answer as output tokens. At any real query volume, the LLM is almost always the bill.
The lever most teams miss: retrieval quality. Sending 6 precise chunks instead of 15 mediocre ones cuts input tokens — and improves answers — more than swapping to a cheaper model. That's why CiteKit ships an eval harness: cheaper and better.