An air-gapped AI deployment runs with no connection to any external network — no internet, no vendor, no model API — and still delivers citation-backed search and chat over your own content. For the most sensitive environments, this is the strongest form of the sovereignty promise: there is no outbound path at all, so there is nothing to leak and nothing to compel.
What “air-gapped” really means
Air-gapping is physical or logical isolation: the system has no route to the outside. It is the posture used for classified work, critical infrastructure (KRITIS), and the most regulated corners of finance, defence and the public sector. The hard requirement that follows is simple to state and easy to fail: everything the platform needs must run inside the boundary. Any hidden dependency — a licence check, a model API, a telemetry beacon — breaks the deployment the moment the network is cut.
The two things that usually break offline
How RAGSuite runs fully offline
Every part of the loop is inside the boundary
- Local models via Ollama — embedding and generation inside your network, no external inference call.
- Offline licensing — Enterprise unlocks with a signed token validated locally; no licence server to reach.
- No phone-home — no telemetry, no usage beacon, no remote kill-switch.
- Self-contained stack — PostgreSQL, Redis and the vector store all live inside the boundary.
The test that proves it: cut the network, then run AI Search and the AI Assistant over your documents. They keep answering, with citations, because every part of the loop is inside with you.
When you actually need it
Air-gapping is overhead, and not every team needs it. It earns its cost when isolation is a hard requirement. Many enterprises run a connected deployment for most projects and reserve air-gapped instances for the few that truly require it; the same platform supports both. See self-hosting architecture and operations and the sovereignty page for the surrounding detail.
Frequently asked questions
When do we actually need air-gapping?
When isolation is a hard requirement — classified or export-controlled data, KRITIS operations, or a security posture that forbids any outbound path. Many enterprises run a connected deployment for most projects and reserve air-gapped instances for the few that require it; the same platform supports both.
How do we prove it's really isolated?
Cut the network, then run AI Search and the AI Assistant over your documents. They keep answering, with citations, because every part of the loop — retrieval, generation via Ollama, and licensing — is inside the boundary with you.
Sources & further reading
- Ollama — run open models locally — local inference, no external API
- Self-hosting enterprise AI: architecture and operations — the surrounding stack