Ollama email sorting combines a local model with a mail-processing application. Running Ollama on your laptop does not make the hosted AI Email Filter service local: the application still accesses your mailbox, and its server must be able to reach the model.
First decide where the application runs
| Application | Model | What to check |
|---|---|---|
| Hosted AI Email Filter | Your own endpoint | Reachability, processing by the hosted app and the configured fallback |
| Self-hosted application | Local Ollama model | Network policy, outbound requests and fallback configuration |
| Either application | Ollama cloud model | The model request can leave your infrastructure |
The current production URL guard rejects loopback and private addresses in custom endpoint settings. A laptop’s localhost cannot be reached from the hosted service. Do not expose an unauthenticated Ollama server publicly to work around this. A private production integration needs an administrator-reviewed deployment configuration.
Check Ollama separately before connecting mail
Install the version for your operating system from Ollama’s official download page. Pick a local model that fits your hardware; this article does not establish a winning model or a minimum RAM requirement.
ollama pull llama3.2:3bollama listcurl http://localhost:11434/api/tagsThe model tag is an example, not an accuracy recommendation. Check that it is a downloaded local model. Ollama also offers cloud models, so a local API address alone is insufficient proof of local inference.
For an allowed deployment, the OpenAI-compatible base URL is http://localhost:11434/v1 only when the application can reach that loopback interface. Use the exact downloaded model tag. Local Ollama does not validate an API-key placeholder; a remote gateway may require real authentication. See the OpenAI compatibility documentation.
Check fallback before claiming local-only processing
In the current application code, a custom endpoint can have the configured default model provider attached as a fallback. If the custom call fails, classification may be attempted there. Selecting your own model is therefore not a local-only guarantee.
Before using confidential messages, verify the deployment’s effective provider chain and test a model outage with synthetic mail. Confirm from network requests that no external endpoint receives the test payload. We have not performed that end-to-end Ollama outage test for this article.
Evaluate sorting with a review set
Start with synthetic invoices, promotions and ambiguous messages. Write down the expected category, run a dry preview and compare each result. Use folder moves first. Permanent deletion cannot be undone through History, and a forwarded message cannot be recalled.
The classifier uses sender, subject and a body preview; it does not automatically know your customers or purchase history. A saved sender category is also conditional rather than a permanent decision for all future messages.
Continue with Bring Your Own LLM and self-hosted email sorting for the application’s documented boundaries.