Skip to main content

Metrics

The admin panel provides three dedicated metrics tabs for monitoring different aspects of platform activity: Model Metrics for AI model performance, Event Metrics for system-wide event tracking, and Slack Metrics for Slack integration analytics. Each tab offers filtering, visualization, and data exploration capabilities.

Model Metrics​

The Model Metrics tab provides detailed analytics on AI model usage and performance across the platform. It is accessible from the General Ops section of the admin sidebar.

Filters​

The filter bar at the top of Model Metrics applies each filter as you pick it. There is no Apply button. Period, Model and Status are kept in the page URL, so a filtered view can be bookmarked or shared; the User filter is not.

FilterDescription
PeriodOne of Today, Last 24 hours, Last 7 days or Last 30 days. Overview, Raw data and the CSV export show at most the 500 most recent matching calls; Analytics runs its own query with a higher cap.
ModelFilter by AI model
UserFilter by specific user
StatusDone, Error or Pending

A Short names / Full names control next to the filters switches between simplified model names and full identifiers. When a filter matches nothing, use Clear filters to reset them all.

Tabs​

Model Metrics is organized into four sub-tabs:

TabDescription
OverviewSummary cards and charts showing key metrics and distributions for the filtered dataset
AnalyticsDetailed chart visualizations with trend analysis based on the applied filters
Raw dataA sortable table of individual metric records with all fields visible. Supports column sorting and includes access to performance metrics documentation via an info modal.
SpendSpend for the selected period, user and model

CSV Export​

Click the Export CSV button in the page header to download the current filtered and sorted dataset as a CSV file. The filename includes a timestamp (e.g., model-metrics-2026-02-15-14-30-00.csv).

Performance Metrics Info​

The Raw data tab includes an info modal that documents performance metrics fields, including streaming performance data (chunk count, total stream time) when available in the dataset.

Time to First Visible Token (TTFVT)​

TTFVT measures how long a user waited before the transcript showed them anything. It counts only text that actually renders: an extended-thinking model's hidden reasoning does not stop the clock, and neither do the tool-call arguments a model writes before answering.

A turn that streamed but never rendered visible text -- reasoning-only, or one that errored before answering -- has no TTFVT rather than a fast-looking one. The Raw data tab shows those as never rendered, alongside the time the first chunk of any kind arrived, which tells you the model did respond while the user saw nothing.

Turns recorded before this measurement was corrected keep the old meaning, and there is no way to repair them: the data needed to recompute them was never stored. They are not distinguishable in this dashboard -- they simply show a TTFVT measured under the old rule. Expect a one-time step in the First Token trend when the fix ships: thinking-heavy models rise toward their true latency, and turns that never rendered drop out of the average entirely. That step is the measurement changing, not the product getting slower, so do not compare across it.

Troubleshooting: TTFVT looks better than users report​

  • never rendered rows in Raw data. These are frozen turns, and they are the ones users complain about. They are excluded from the First Token trend chart rather than averaged into it, so the chart cannot show them -- scan the Raw data tab instead, where they are flagged in red. There is no TTFVT sort key, so scanning is the only way in. There is no dedicated filter or counter for them yet; more of them appearing is the signal, even while the chart looks flat.

  • TTFVT far below Client First Token Time. Both apply the same visibility rule, so ordinarily TTFVT stops at the server and the gap is network plus client render time. Compare the two before attributing slowness to the model.

    Three cases break that comparison, and all three make the client number the odd one out. On an append-mode rapid reply the server discounts the seeded prefix and the client does not. On a retried turn the server's stamp is relative to the attempt that finally answered, while the client records once per quest and so stays relative to the first attempt. And a turn with no TTFVT at all may have been retried rather than never streamed. Treat a large gap on any of these as an artifact of what each side measured, not as client-side slowness.

  • Turns with no TTFVT and no first-chunk time. The response was not streamed at all. Check the model's streaming support rather than reading it as a latency problem.

Record Count​

The page header shows how many calls match the current filters. The count is hidden once a query reaches the 500-call cap, because it would then be a floor rather than a total.

Event Metrics​

The Event Metrics tab tracks system-wide events across the platform, providing visibility into user actions, system events, and integration activity.

Filters​

FilterDescription
Date RangeQuick ranges 7d, 30d, 90d and All time, plus a Range chip with Today, Yesterday, This month, Last month and Custom (whole From and To days)
UserFilter by specific user
EventFilter by event type
CategoryFilter by event category (e.g., Slack, System, User)

A filter applies as soon as you pick it. Clear filters resets the range to All time and removes the other filters.

Tabs​

Event Metrics is organized into two sub-tabs:

TabDescription
OverviewSummary visualizations and charts for the filtered event data
Curation breakdownDetailed breakdown of events by curation category, showing distribution patterns

When no curation events were recorded, the Curation breakdown tab says No curation events yet with a Refresh metrics button. When a range or filter is set, it says No curation events in this range with a Clear filters button instead.

Key numbers​

A row of key numbers above the tabs shows Total events, Filtered, Unique events and Unique users for the current filters.

By user​

The Events / By user switch at the top of the tab opens the per-user activity log: one row per day, action and user, with a metadata sample and a count. The view is kept in the page URL.

ControlDescription
OrganizationsLimit rows to one or more organizations
ExcludeHide rows without an organization, and staff and test accounts
Advanced filtersDate range presets, action and email search, and metadata filters
Export CSVDownload every row that matches the filters, not just the current page

Slack Metrics​

The Slack Metrics tab provides focused analytics for Slack integration usage. It reuses the Event Metrics filtering infrastructure but automatically restricts the data to the "Slack" event category (the category filter selector is hidden in the UI).

Filters​

FilterDescription
Date RangeQuick ranges 7d, 30d, 90d and All time, plus a Range chip with Today, Yesterday, This month, Last month and Custom (whole From and To days)
UserFilter by specific user
EventFilter by specific Slack event type

Visualizations​

The Slack Metrics tab displays several chart panels:

ChartDescription
Slack Event ActivityHorizontal bar chart showing event counts by type
Agent UsageDonut/pie chart showing distribution of agent persona usage in Slack interactions
Intent BreakdownDonut/pie chart showing the distribution of detected intents
Export FormatsDonut/pie chart showing which export formats (JSON, CSV, Markdown) are used
Export Success RatePie chart showing the ratio of successful to failed exports

When no Slack data is available, the tab says No Slack events yet with a Refresh metrics button. When a range or filter is set, it says No events match these filters with a Clear filters button instead.

Error Handling​

If the metrics data fails to load, an error message is shown with a Try again button that loads the data again.

Best Practices​

  • Use date range filters to scope queries to relevant time periods, which improves load times and chart readability.
  • Set filters before switching between sub-tabs -- the filtered dataset persists across tabs within each metrics view.
  • Use the CSV export in Model Metrics to perform offline analysis or share data with stakeholders.
  • Monitor the Slack Metrics export success rate to identify integration reliability issues.
  • Check the Event Metrics Curation breakdown to understand content categorization patterns.