=

The passage around this formula

What the convergence does support is a real, worked comparison. Write the retrieval pipeline’s monthly cost, using the terms established above, as a fixed storage component plus a per-query component that itself has two parts: $C_{\mathrm{RAG}}(Q) = p_{\mathrm{store}} \cdot S \;+\; Q \cdot \left( p_{\mathrm{ru}} \cdot r(S) + p_{\mathrm{rerank}} \right)$. where S is index size in gigabytes, r(S) is Pinecone’s documented read units per query as a function of that size, Q is monthly query volume, and $p_{\mathrm{store}}$ , $p_{\mathrm{ru}}$ , and $p_{\mathrm{rerank}}$ are the published per-unit rates given earlier. Set against it, the cost of skipping retrieval and instead inlining a fixed block of extra context on every single call is

Read this part in the article →