Anvit: Local Agentic RAG
100% On-Device AIAnvit is a privacy-first AI assistant that runs entirely on your device. It lets you chat with your documents using a local Gemma 4 language model โ no cloud inference, no analytics, no tracking.
This policy explains what data Anvit stores locally, what limited information may leave your device, and how you remain in control at all times.
All of the following is stored only on your device and is never transmitted to any server:
| Data | Purpose |
|---|---|
| Chat history & messages | Persist conversations across sessions |
| Uploaded PDF documents | Source material for document Q&A |
| Text embeddings (vectors) | Enable local semantic search (RAG) |
| AI model files (LLM & embedding) | Run inference fully offline after download |
| App settings & preferences | Remember your configuration |
| HuggingFace token Optional | Authenticate model downloads (if provided) |
| Email address Optional | Pre-fill the feedback form (if provided) |
Anvit requests the RECORD_AUDIO permission to enable voice-to-text interactions with your local AI agent.
On-Device Processing: All audio recording is processed strictly on your physical device using local speech-to-text engines.
No Cloud Transmission: Audio data is never uploaded to any external servers, third parties, or cloud providers.
No Retention: Raw audio files are deleted immediately after the local AI agent completes the transcription process.
Anvit makes network requests in two narrow, user-initiated situations:
When you download a language or embedding model in Settings, Anvit fetches model files from huggingface.co. No personal data is sent. If you provide a HuggingFace token it is used only as an authorization header for faster downloads and is stored locally.
If you tap Report inaccuracy and submit the form, the following is sent to Google Forms:
This is entirely optional. You can dismiss the feedback dialog at any time without sending anything.
No other network calls are made. All AI inference, document retrieval, and embedding computation runs locally. Anvit has no telemetry, crash reporting, or background analytics.
| Permission | Why It's Needed |
|---|---|
INTERNET |
One-time model downloads from HuggingFace; optional feedback submission. Not used for inference. |
RECORD_AUDIO |
Voice-to-text queries. Audio never leaves your device. |
| File / Storage access | Selecting PDF files via the system file picker (scoped storage โ only files you choose). |
Anvit does not knowingly collect any personal information from children under 13 (or the applicable age of digital consent in your jurisdiction). The app is intended for general audiences. If you believe a child has provided personal data through the voluntary feedback form, contact us and we will remove it.
Anvit interacts with the following third-party services in limited, user-initiated contexts:
No other third-party SDKs, analytics libraries, or advertising networks are included in the app.
Because all data is stored locally on your device, you can delete it at any time:
For feedback submissions already sent via Google Forms, contact us to request removal.
If we make material changes to this privacy policy, we will update the effective date above and โ where feasible โ notify users within the app. Continued use of Anvit after changes are posted constitutes acceptance of the updated policy.
Questions or concerns about this policy? Reach out at: likhithv02@gmail.com