How to keep confidential data fully local

Goal

Use private or confidential data so that it does not leave the machine — or, in mixed operation, so that your private vault stays local while you pair with a central instance for shared knowledge. The model behind this (three levels) is explained in the confidentiality model.

Variant A — Maximum confidentiality (nothing leaves the machine)

Recommended for highly confidential data.

Prerequisites: the desktop app; a local model (e.g. Ollama/LM Studio) on the same machine.

1. Run the desktop autonomously — no pairing with a central instance. → Single-seat minimal (autonomous desktop)
2. Set up a local model — create a model provider of type Ollama/LM Studio with a local endpoint (e.g. http://localhost:11434). → Admin tab: AI models / providers
3. Choose sqlite-vec as the vector backend — so the search vectors stay local (no Qdrant). → Set up Qdrant (for contrast) · RAG & search
4. Create a private vault — personal/local, with appropriate access rights. → Create a vault and assign access rights
5. Add data — import documents or store chat answers in the private vault. → Import a document

Result: content, vectors and model answers all stay on the machine.

Variant B — Convenience: a private vault alongside a central instance

Recommended when you want to use shared knowledge but keep confidential material local.

Prerequisites: the desktop app; access to a central instance / a peer server.

1. Pair with the central instance (server mode/client). → Peer sync / server mode in operation
2. Create a private local vault — separate from the shared vaults. → Create a vault
3. Put confidential data into the private vault only; shared knowledge into the shared vaults.

Result: the private vault (content + vectors) stays local and is never synced; shared vaults are searched live on the server, and the search fuses both.

Verification

• Variant A: in the model tab only the local model is active; the vector backend shows sqlite-vec (or “Local index”). A search in the private vault returns matches with no internet connection.
• Variant B: the private vault does not appear in the configuration sync (the connections diff shows settings only). Shared matches appear additionally.

Important limits (named honestly)

• Cloud model: if you use a cloud model instead of a local one, input and RAG context are pseudonymized but still transmitted — and some things deliberately go out in plaintext, such as source references and workspace files. “Nothing leaves the machine” only holds with a local model. → Keeping confidential data local — the model boundary
• The backend is global: a desktop uses one vector backend for all local vaults — not selectable per vault.
• Backup: a backup of the private vault is encrypted; keep the passphrase safe.

Related

• Keeping confidential data local — the confidentiality model
• Data sovereignty & dedicated instance · Single-seat minimal (autonomous desktop)