Glossary
Private AI
AI that works on your data without handing that data to someone else’s model service. Private AI covers everything from a zero-retention API contract to a fully air-gapped model on your own hardware.
General definition
Private AI is any deployment of AI models and agents where the data you send in, the context retrieved for the model, and the answers that come back are processed inside a boundary your organisation controls, or under contractual terms that stop a third party from keeping or training on them. It is a goal, not a single architecture. Teams reach it in several ways, and the right one depends on how sensitive the data is, which regulations apply, and how much infrastructure you are willing to run.
The options form a spectrum from least to most isolated:
- Zero-retention public API: a cloud model provider agrees in writing not to store prompts or outputs or use them for training (see zero data retention). Data still leaves your network.
- Private endpoint on a public cloud: the model runs in the hyperscaler’s managed service but is reached over a private link from your virtual network, in a region you choose, under the cloud provider’s data terms.
- Dedicated or self-hosted model in your cloud account: an open-weight model served by vLLM or Ollama on GPUs you rent, inside your own VPC. This is the private LLM pattern.
- On-premise: the same stack on hardware in your data centre, with controlled egress.
- Air-gapped: no network path out at all; models and updates arrive on physical media.
Each step to the right buys more control and more operating work. A zero-retention API needs a contract review; a self-hosted model needs GPU capacity, patching, model updates and someone on call. Private AI is also wider than the model itself: the chat application, the vector store used for retrieval, logs, and any embeddings pipeline all handle the same sensitive content and need the same treatment. Regulators care about where personal data is processed (data residency) and who can access it, so the honest question is not “is the model private” but “where does every copy of the data go”.
In the Ethora ecosystem
Ethora is a chat and AI agent platform that you can run as managed cloud, as a dedicated server inside your own cloud account, or fully self-hosted with Docker Compose or Helm. In the self-hosted and dedicated options, messages, files, users, the RAG knowledge base and the audit trail live on infrastructure you control. The AI service talks to any OpenAI-compatible endpoint, so each deployment picks its own point on the spectrum: a cloud model API, a private endpoint in your tenancy, or a local model served by Ollama or vLLM on the same network.
That makes private AI a configuration decision rather than a rewrite. The self-hosted LLM AI agent page describes the local-model route, including the official Ollama example where switching to a local model is a handful of environment variables. Ethora does not certify your deployment as private; it gives you an open-source stack with no built-in dependency on Ethora-operated services once installed, which is what an isolation review needs to see.