RAG over the company wiki, with embeddings computed on the box

Mount the wiki export into a container on your bare metal server. Run an embedding model and a chat model as daemon programs on localhost. A script walks the pages, embeds each one, and writes the vectors into a SQLite file under /hoody/databases/, where the host-level mount keeps the writer and the query path from colliding. hoody-agent points at the local chat model and answers by reading that database.

Nothing went to an embeddings provider. Hoody's API sees a container and a snapshot; it doesn't see the wiki or the vectors.

The wiki, its vectors, and the model that reads them share one disk in one rack.