Troubleshooting

AI features are unavailable

With the in-app backend, check that the recommended model finished downloading in Settings → AI Backend & Models. An interrupted download leaves the model unusable — remove it and download it again.

If you switched to Ollama, make sure the daemon is running and reachable at http://localhost:11434, and that you have pulled a model:

ollama pull llama3.2
ollama pull nomic-embed-text

On an Intel Mac, the embedded runtime is not included in the build. Use Ollama or OpenRouter instead.

Chat is slow

Local inference takes real time — on a modest machine, a chat answer can take tens of seconds. Things that help, in rough order of effect:

1. Check the model actually fits. This is the big one. On Windows or Linux, a model larger than your GPU’s VRAM spills over to the CPU and slows down by several times — the fix is a smaller model, not more system RAM. On Apple Silicon the comparison is against total unified memory. See the model catalog for the figure per model.
2. Use a smaller model. Qwen 3.5 4B is the recommended default for a reason.
3. Raise “keep model loaded” in AI settings so the model is not reloaded from disk on every question.
4. Lower the context window — a smaller window means less to process per turn, and it is what to reduce first when a model only just fits.
5. Turn off thinking mode, which trades speed for accuracy.

The GPU is not being used (Windows / Linux)

The app log says which device a model was loaded onto. A working GPU load looks like:

llamacpp: chat model offload — Vulkan0 (Vulkan) has 15 GB free — offloading all layers

If you do not see a line like that, the Vulkan backend found no usable device and quietly fell back to the CPU — the app still works, just slower. Check, in order:

1. Your graphics driver. This is almost always the cause. Install or update the normal driver for your card; no CUDA toolkit or vendor SDK is needed.
2. That Vulkan sees the device. Run vulkaninfo --summary (from vulkan-tools). If it reports no device, the problem is below EmailOps — fix the driver stack first.
3. VRAM headroom. If the log offloads only some layers, the model is bigger than the card’s free VRAM. Pick a smaller model or lower the context window.

Virtual machines and remote desktops frequently expose no GPU at all, which is expected.

Search returns keyword results only

Semantic search needs embeddings. Open Settings → AI Search, check that the categories you care about are selected, and let the embedding pass finish. After changing the embedding model, rebuild the index from the same screen.

Also check Limit AI processing to recent emails in AI settings — mail older than that window is skipped deliberately.

Classification is not tagging anything

• Confirm auto-classify new emails is on in Settings → AI Classification.
• Check which Gmail categories are selected; if none are, nothing gets classified.
• For mail that arrived before you enabled it, use Classify Unclassified, or Reclassify All after changing the prompt or rules.

Gmail sync stalls or reports rate limits

Gmail enforces per-account quotas. When it asks EmailOps to back off, sync pauses that account until the window reopens and resumes on the next scheduled run — no action needed. If sync stays broken, remove and re-add the account so a fresh token is issued.

The app is locked and I forgot the main password

The main password is a local lock with no recovery path — that is the point. Your mail is still on the server; you can reinstall EmailOps against a fresh data directory and re-sync.

Something else

Check the open issues and, if your problem is not there, open a new one. Include your OS and version, the EmailOps version, which AI backend and model you use, and what you expected to happen.