Michael Limberger
Need me? Email mike@limberger.ca
AI
Troubleshooting
When the window won't open
Things will break.
It is software.
If localhost:3000 gives a connection error, run docker compose ps.
If OpenWebUI is not listed, run docker compose up -d.
If it is in a restarting loop, run docker compose logs --tail 50.
If port 3000 is already in use, change the compose mapping to 8080:8080, bring the stack down and up, and browse http://localhost:8080.
Empty model dropdown usually means OpenWebUI cannot reach Ollama.
Check the menu bar icon, or run curl http://localhost:11434/api/version.
From Docker, the URL should be http://host.docker.internal:11434.
Set it under Admin Panel, Settings, Connections, Ollama, then verify.
After that, run ollama list.
If nothing is there, pull one: ollama pull qwen3:30b-a3b.
Gibberish responses often mean temperature is above 1.5.
Lower it to about 0.7.
Mirostat and Top P together can also get weird.
If a long system prompt is eating the window, shorten the prompt or raise context.
Extremely slow chats: open Activity Monitor and check memory pressure.
Red means swap.
Use a smaller model, close Safari tabs, and run ollama ps.
PROCESSOR should show GPU, not CPU.
If it forgets earlier context, the context window was exceeded.
Start a new conversation per topic, raise num_ctx (it costs RAM), keep messages shorter, or ask for a summary and carry that forward.
If an image is ignored: is a vision model selected?
Is the custom model's Vision toggle on?
Is the thumbnail visible before you send?
Huge images can fail silently, so try a smaller crop.
Docker eating disk: docker system prune -a removes unused images.
The open-webui volume stays.
To update: docker compose pull then docker compose up -d.
Data persists across that update.
Nuclear option: docker compose down -v wipes conversations, settings, and accounts.
Help: github.com/open-webui/open-webui, Discord linked from the repo, docs.openwebui.com, or the next mixer. Next: the card you keep nearby.