VPS for self-hosted RAG at scale

A RAG demo on a laptop with two hundred documents feels effortless. Then you point it at a real corpus — a company's docs, years of tickets, an actual knowledge base — and the whole thing becomes a memory problem. Not a CPU problem. A memory one.

Here's the part most hosting pages skip: fast retrieval wants the index in RAM. Every chunk of text becomes an embedding — a vector a few hundred to a couple thousand numbers wide — and searching means comparing your query against all of them, quickly. On disk it works, but every query pays a latency tax, and low-latency retrieval was the reason you self-hosted in the first place.

The sizing math, honestly

Measure your own corpus — dimension and index type swing this a lot — but as a starting feel:

• A few hundred thousand embeddings — 2–4 GB. A personal knowledge base. You don't need Pro for this; a smaller plan is fine.
• Low millions — with the app, the model client and the OS around it, plan for 16–32 GB. A serious company knowledge base. This is where Pro-32 or Pro-48 starts making sense.
• Tens of millions, or high-dimension vectors — 48 GB and up, and past ~80 GB you're splitting the index across servers. Large document estates, multi-tenant retrieval, several indexes hot at once.

We wrote out the full RAM curve here if you want the details.

Why high-memory and no-KYC together

Renting 48 GB of RAM is easy. Renting it with crypto and no identity check is not — most hosts that sell serious memory cheaply do it behind a card and a KYC form. Your embeddings aren't abstract numbers; they encode the text they came from. Your docs, your customers' content, turned into vectors. If that data is sensitive enough that you're paying privately, a managed vector cloud undoes the whole point — and so does a host that ties the server to your identity.

That combination — high memory, dedicated IP, crypto, no KYC, nightly backups — is what the Pro line is for. It's not the cheapest per gigabyte, and if you don't need privacy you can find RAM cheaper elsewhere. But if the index is your product and it can't leave, this is the shape that fits.

Where to start

Pick a plan with headroom for the index plus everything around it — the app, the model client, room to grow. For most real corpora that's Pro-48 (48 GB); a large one goes to Pro-64 or Pro-80. Load a sample first, watch memory, size from the number you measured — not the one you feared.

If your retrieval is part of a bigger agent system, the same box often holds the agent fleet too — that's how a 48 GB plan quietly becomes a 64 GB one.