Privacy architecture

Local by design. Permission by exception.

The assistant runs on your computer, rejects remote model configuration and keeps every sensitive collection outside the public website and source history.

Stays local

Your private material

  • Chats and project history
  • Approved memories
  • Knowledge Vault documents
  • Embeddings and indexes
  • Dataset contents
  • Repository previews
  • Generated artifacts
  • Private evaluation answers
Explicit permission

Local access boundaries

  • Repository allowlisting
  • Dataset allowlisting
  • Per-chat attachments
  • On-demand file previews
  • Memory approval
  • Reversible access
  • Visible calculation receipts
  • Local backup and deletion
Locked by default

External paths

  • Cloud model handoff
  • Automatic web browsing
  • Email access
  • Calendar access
  • Telemetry
  • Remote Ollama servers
  • Silent uploads
  • Public model fallback
Honest threat model

Local does not mean invulnerable.

Malware or another user with host access may read local files. Models can still produce incorrect content. Approved repositories and datasets may contain sensitive information. Rangabot reduces network exposure; it cannot replace operating-system security or human review.

01

Loopback model

Non-loopback Ollama URLs are rejected so configuration mistakes do not silently transmit prompts.

02

Private storage

Databases, vault material, approvals and artifacts are ignored by Git and omitted from website builds.

03

Visible uncertainty

Unavailable live data and unsupported actions must be stated rather than fabricated.

Version 1.2 local state

Recovery and context remain explicit.

The release adds stronger local state handling without expanding external access.

01

Manual recovery

A failed turn may retain the exact local request, response mode and reply receipt for review. Failed partial output is not reused, and nothing is sent automatically.

02

Revisioned data meaning

Dataset descriptions, aliases and confirmed joins are saved locally against the exact dataset identity. They cannot grant file access or add a requested operation.

03

Profiles are organization

Profiles separate application-managed workspaces. They do not isolate the wider filesystem, add encryption or promise secure deletion.