Quick diagnosis
An app installed on your computer can still send prompts or documents to a remote model, embedding service, web tool, or synced folder. Check the exact workflow with a harmless sample before entering sensitive material.
First three checks
- Inspect the selected chat model and provider address; distinguish a downloaded local model from a cloud model or remote endpoint.
- Check the embedding, document parsing, web search, and connected-tool settings separately from the chat model.
- Find the chat, document, index, and log storage locations and check whether backup or sync software copies them elsewhere.
Check the model and every helper
In Open WebUI or AnythingLLM, open the provider settings and inspect the selected model endpoint. A local Ollama address may look like http://localhost:11434; in Docker, the address can instead identify the host or another container. Establish which computer that address actually reaches.
In Ollama, use an explicitly downloaded local model with ollama run MODEL_NAME. A cloud-tagged model or a provider chosen in an agent is a different data path. In LM Studio, distinguish local desktop inference from a linked computer or a hosted execution choice. A vendor name alone does not identify where a task runs.
For document chat, inspect the embedding provider too. A local chat model paired with remote embeddings can send document text away during indexing, before the first question. Also check OCR, parsing, reranking, web search, and enabled plugins or MCP servers.
Check storage and network access
| Boundary | Concrete check | Action if it is wrong |
|---|---|---|
| Stored documents and indexes | Locate the actual app/workspace folders | Move or configure storage using the app's supported controls if it is in an unwanted sync location |
| Chats and logs | Inspect retention, exported transcripts, and debug logging | Reduce unnecessary retention and avoid logging sensitive sample prompts |
| Remote helpers | Inspect each enabled tool and provider | Disable the helper or deliberately choose an acceptable data destination |
| Model API | Inspect the service's listening address and firewall scope | Keep access limited to the intended machine or authenticated private network |
| Browser interface | Check whether other devices can reach its published port | Restrict access rather than disabling authentication |
An address such as 0.0.0.0 means the service listens on all available interfaces; it is not an address that proves private access. If another machine needs the service, plan authentication and firewall access explicitly. See the Open WebUI connection guide for container-specific addresses.
Test the intended offline workflow
Finish downloading the model and required runtimes. Use a made-up document, disable optional remote features, disconnect outbound networking, and repeat both document import and chat. Confirm that the intended tasks still work and inspect failures instead of silently switching to another provider.
If plain chat works but importing the sample fails, investigate embeddings, OCR, or another ingestion dependency. If only web search fails, that feature requires a network path. An offline success checks that sample workflow; it does not establish what the application sends when networking is restored or what another feature does.
When to stop
Pause sensitive input when a provider, helper, storage location, or network listener remains unexplained. Resume after the unwanted data path is removed or intentionally accepted. Use the full local AI privacy guide and PDF setup guide to check the complete workflow.