Performance baseline
- Start/stop debounce:
2000ms. - Manual answer trigger debounce:
300ms. - Auto-queue trigger debounce:
500ms. - Rule out debounce windows first, then decide whether the response is truly slow.
Link characteristics
- Answers are returned as streams. A new request actively aborts the old request.
- "The previous answer suddenly stopped and the next one started outputting" is often request switching, not necessarily a backend failure.
- A 200 response with an empty text stream can happen. It means this request produced no valid content.
Quick diagnosis steps
- Establish a minimum baseline: turn off proxy/VPN, keep the page in the foreground, and temporarily disable the knowledge base.
- Run three short sessions in a row and record timing and behavior for each.
- Restore configuration one variable at a time: model, knowledge base, extra requirements, and observe changes.
- Attribute by symptom:
- Transcript is stable but answer is slow: check model and knowledge base first.
- Speed varies greatly: check network and browser throttling first.
- Occasional no output: check trigger conditions and whether the request was aborted.
Escalate to support
- Timeouts continue across network environments.
- Rebuilding the session does not recover.
- Other accounts on the same device work normally, but the current account remains abnormal.
When contacting support, include browser version, network environment, reproduction steps, and key response codes.