Where does your data go when you use AI Box?

“Local-first” describes a deployment choice, not an automatic guarantee that every request stays on one device. The model endpoint, network configuration, optional integrations, and website services all matter. Review those choices before indexing confidential material.

The document path

1. A user uploads a supported file to the AI Box application over the configured network.
2. The backend extracts text, divides it into searchable passages, and creates embeddings using the configured embedding model.
3. Document metadata, extracted passages, and embeddings are stored in the appliance database for retrieval.
4. When a user asks a question, the system retrieves relevant passages and sends the question plus retrieved context to the configured language-model endpoint.
5. The answer can include source document and page references; users should verify those references against the original file.

The model endpoint changes the boundary

AI Box supports local model configurations, including a local GGUF model or an Ollama-compatible endpoint. It can also be configured to use external AI providers. When a remote endpoint is selected, the prompt and retrieved document passages may be sent there. A custom endpoint can also point to another server on your network, so confirm where it actually runs.

• Confirm the active model provider and endpoint with the system administrator.
• Check firewall and DNS logs from a representative deployment rather than relying on the product label alone.
• Test the workflow with internet access blocked after setup if offline operation is a requirement.
• Treat the website support chat separately: it uses a hosted AI service and is not the appliance's local inference path. Do not paste confidential documents into it.