RAG System Cost Calculator

This calculator walks through a RAG pipeline's cost stack in the order money actually gets spent. First, chunking: your document count and average tokens-per-document get split into overlapping chunks of the size you specify, using (tokens − overlap) / (chunk size − overlap) to find how many chunks each document yields — larger overlap means more chunks (and more redundant embedding cost) for the same document, since consecutive chunks re-embed a shared slice of text to preserve context across boundaries. Every one of those chunks gets a one-time embedding cost, calculated here at OpenAI's ada-002 rate of $0.10 per million tokens regardless of what embedding-dimensions value you enter — so if you set 3072 dimensions to model text-embedding-3-large, the storage math correctly grows (each vector costs dimensions × 4 bytes for float32 plus 50 bytes of metadata overhead) but the embedding-cost line still assumes ada-002 pricing, a real limitation to keep in mind when comparing providers.

Vector database hosting is approximated linearly at $0.07/month per 1,000 stored vectors, modeled on Pinecone's roughly $70/month per million vectors. Ongoing query cost has two parts: embedding each incoming query (same ada-002 rate) and an LLM synthesis call sized by Top-K retrieved chunks × chunk size as input tokens plus a fixed 500 output tokens, priced at GPT-4o's published $2.50/1M input and $10/1M output rates. Because the embedding and LLM pricing are both hard-coded to specific providers, use this as a directional budget for a given architecture, then substitute your actual provider's rates for a precise number.