Model Quality Monitoring Q&A

On the Service Quality Monitoring scatter plot page, why does the diagnostic log display two "Total Token Count" values? How is the billable token count calculated?
WeCom screenshot

Answer: To validate the Cache Hit Rate and Repeat Call Stability metrics, the system sends the same question to the model twice, producing two sets of metric results. The report uses only the result from the first call, while the billable token count is the sum of the token counts returned by both calls.

For model output quality monitoring tasks, why does a timeout trigger when the task timeout is set to 300 s but the actual total download time reaches only 100 s?
WeCom screenshot

Answer: The 300 s timeout is evenly allocated across each step, so each step has a 100 s timeout.

For output quality tasks, why is the model response content not fully displayed?
WeCom screenshot

Answer: For overly long output results, the content is truncated to a maximum of 2,000 characters.

What is the retention period for performance (scatter plot) data in Model Quality Monitoring?
Answer: The retention period is 3 months. Long-term data retention is not available.