Troubleshooting¶

Start at the top and follow the arrows.

flowchart TD
    S([Something is wrong]) --> Q1{What do you see?}
    Q1 -->|"Red box: 'did not return<br/>a usable answer'"| A1["Turn Thinking mode OFF<br/>or switch model"]
    Q1 -->|"'No recent news found'"| A2["Check Sources page,<br/>network and the ticker"]
    Q1 -->|"Stuck on 'Waiting for<br/>the model'"| A3["Another run holds the model:<br/>wait, or press Stop there"]
    Q1 -->|"Very slow, tens of seconds<br/>or minutes"| A4{Which model?}
    A4 -->|"cloud model"| A5["Parallel requests queue.<br/>Run one at a time"]
    A4 -->|"local qwen3"| A6["Thinking OFF,<br/>or use a cloud model"]
    Q1 -->|"503 'Add an OpenAI API key'"| A7["Settings: paste the key"]
    Q1 -->|"Settings change ignored"| A8["Env vars only seed.<br/>Use the Settings page"]
    Q1 -->|"Page loads, API calls fail"| A9["Is the API on :8000?<br/>See ports below"]
    Q1 -->|"Some sources failed: ..."| A10["Read the reason, then use<br/>Test feed on the Sources page"]

Symptoms¶

"model did not return a usable answer"¶

The model replied to every attempt (up to 4 calls) with nothing, or with JSON that does not match the schema. Typical for local reasoning models (qwen3), which can spend the whole turn thinking.

1. Settings, turn Thinking mode off, Save, try again.
2. Still failing? Use a bigger or cloud model (gpt-oss:120b-cloud).
3. The Settings page Token usage shows how many requests a run needed; more than 1 means it retried.

Very slow runs¶

The status line shows live elapsed time; the final badge shows the total.

"Waiting for the model"¶

Only one model call runs at a time. Another tab, or another person, is using it. Wait, or open that run and press Stop. A run abandoned by closing its tab frees the slot immediately.

I turned Thinking off but still saw reasoning¶

Fixed: with the switch off the server no longer forwards reasoning, even for models that cannot stop thinking (gpt-oss). If you still see it, check the page was reloaded and that you pressed Save (the "Saved" tick appears).

"Some sources failed: ..."¶

The briefing still works from the other sources. The text after the name says why:

Use Test feed on the Sources page to check a URL before saving it. Details and the CNN and Nasdaq examples are in News sources.

Ollama cloud model not found¶

Cloud model names end in -cloud (gpt-oss:120b-cloud), and you must be signed in:

ollama signin
curl -s localhost:11434/v1/chat/completions -d '{"model":"gpt-oss:120b-cloud","messages":[{"role":"user","content":"hi"}]}'

Reading the logs¶

Every request has an id (X-Request-ID response header, request_id on each log line). Find it, then read that request's lines in order: which feed was slow (feed_fetched duration_ms), whether the model had to wait (model_acquired waited_ms), whether it retried (model_retry, retried=True), where a cancelled run stopped (client_disconnected phase=...). Details in Observability.

Ports and processes¶

Starting over¶

rm backend/zeroai.db      # stop the API first: sources, model and usage are re-seeded