Why your search has no meter running

Many AI products work like a courier service: to understand a document, they mail a copy to an outside company that reads it, charges per page, and mails back a summary. Your node does not use a courier. It employs an in-house archivist that reads every document, files it, and answers questions, all without the text leaving the building.

How the archivist turns a document into a searchable fingerprint is the subject of Semantic Search. What matters for cost is where that work happens: entirely in-process, on compute you already own. No document is mailed to an outside reader to be understood, so there is nothing to meter, either when you ingest a document or when you later ask a question about it.

The fingerprints are stored right next to your text, in a vector store that lives in the same deployment, so retrieval is a lookup inside your own boundary rather than a call across it. That is the whole reason there is no meter: nothing about a search has to leave to happen.

Running retrieval in-house collapses two costs at once. The obvious one is money: there are no metered embedding fees that scale with how much you ingest or how often you ask. The less obvious one is exposure. A per-token bill exists because your text crossed a boundary to be counted; removing the meter and removing the egress are the same act.

The retrieval layer is deliberately backend-agnostic: the identical schema and query logic run against the control-plane PostgreSQL pool and against the in-pod PGlite store on a node. A node with no connection to a central database still searches its own content the same way. That portability is what lets an isolated edge node behave exactly like a cloud node without a separate code path.

In-house compute is not free compute; embedding a very large corpus takes CPU time on your node, and larger workloads want larger nodes. The bundled model is small and strongest in English, so retrieval quality tracks that. And "no per-token embedding bill" is a statement about retrieval specifically; any generation model you point your node at has its own separate cost profile.