Embeddings cost CPU minutes, not a per-token fee

Your product indexes each customer's documents for retrieval. Two million chunks at onboarding, re-embedded whenever they change. On a hosted embedding API that's a per-token bill that scales with how much customers upload, which is the wrong thing to be nervous about.

On Hoody a small embedding model runs as a daemon program in a container, served by Ollama from the box's own cores. An exec script walks the customer's documents, calls the model on localhost, and writes the vectors into a SQLite file in /hoody/databases/. The first run takes the night. Re-embedding one changed document takes a second.

The bill for two million chunks is the server's flat monthly rate, which you were paying to host the product anyway. The bill for the next two million is the same number.

The embedding budget is a number of cores times a number of hours, and you already own both.