🛡️ Zero-Knowledge Architecture

Privacy Policy & Data Transparency

Effective Date: September 2026 · Privex AI v2.1.0

Our Core Privacy Commitment

Privex AI is built on the fundamental principle that your thoughts, personal memories, conversations, and uploaded documents belong exclusively to you. All user data is encrypted with AES-GCM-256 via the browser's hardware WebCrypto API before being stored in IndexedDB. Decryption keys never leave your device memory.

1. What We NEVER Store or Transmit

The following data categories are strictly forbidden from leaving your device:

Never Sent Conversation Contents: Every prompt you submit and every response generated is processed and stored locally.
Never Sent Adaptive Memories: Learned preferences, styles, and identity statements stay encrypted on your machine.
Never Sent Knowledge Vault Documents: Uploaded PDFs, notes, and code files are chunked and stored locally.
Never Sent Semantic Vectors & Embeddings: High-dimensional embeddings computed by our web workers are stored exclusively in local IndexedDB.
Never Sent API Keys: Third-party provider keys (OpenAI, Anthropic, Groq, Gemini) are encrypted in browser storage and sent directly to those APIs only when you submit a prompt in Cloud API mode.

2. Anonymous Developer Telemetry

To help improve Privex AI (understanding feature adoption, fleet storage growth, and supported hardware), the app transmits minimal, anonymized diagnostic telemetry to Google Cloud Firestore.

Metric Field	Description	Privacy Safeguard
`sessionId`	Random RFC4122 v4 UUID generated on first app launch	Anonymous Decoupled from personal identifiers
`appVersion`	Installed version (e.g. `2.1.0`)	Diagnostic Fleet upgrade tracking
`storageUsedBytes`	Total IndexedDB ciphertext bytes on disk	Aggregate Pure byte count, no file names
`vaultDocCount` / `vaultChunkCount`	Number of indexed vault documents and chunks	Count No document titles or contents
`memoryCount` / `conversationCount`	Total count of memories and conversation threads	Count Numerical volume only
`activeMode`	Routing mode (`local` or `cloud`)	Preference No prompt or query parameters
`localModelName`	Ollama model identifier (e.g. `llama3.2:3b`)	Model Tag String tag only
`featureUsage`	Booleans: vault, memory, voice STT, charts, PDF export	Boolean Simple true/false usage flags
`hardwareTier`	Coarse profile: `low`, `mid`, or `high`	Coarse No canvas/WebGL fingerprints
`platform`	Client environment (`web` or `desktop`)	Environment Host runtime type

3. Firestore Security & Write-Only Access

Our Google Cloud Firestore database enforces strict security rules:

Write-Only for Clients: Client devices can only submit their own matching sessionId metrics. Clients have zero read permissions across the entire telemetry collection.
Schema Enforcement: The database rejects any write attempt containing prohibited keys (such as conversation, message, prompt, memory, or apiKey).
Developer-Only Auditing: Only authorized project developers can query aggregated telemetry via the Google Cloud / Firebase Admin Console.

4. How to Opt-Out (1-Click Kill Switch)

You maintain complete control over diagnostic telemetry:

Disabling Anonymous Analytics

Open Settings → About → uncheck "Send anonymous analytics". When toggled OFF, all background timers, lazy-loaded SDK calls, and network dispatches are immediately and permanently halted.

Return to Privex AI