Insight
The three screens that answer how the installation is doing, what changed over time, and what is happening in this minute.
Insight is read-only. Nothing on these three screens changes the product, which is what makes it the right place to start a day.
Overview
Insight → Overview is the state of the installation in one page: usage, spend, the mix of models being used,
and a short list of recent activity.
| Area | What it tells you |
|---|---|
| Usage | Turns in the current window, against what the plans allow |
| Spend | Credits granted, spent and in flight, which is the number that costs money |
| Mix | Which models are being used, and in what proportion |
| Recent | The last few things that happened, newest first |
Read it as a dashboard rather than an accounting report: the numbers are rounded to be read at a glance, and the exact figures live one screen along.
Analytics
Insight → Analytics is the same information over time, which is what makes it useful for a decision.
| Series | The question it answers |
|---|---|
| Signups | Is anything bringing people in? |
| Turns | Is the product being used, or is it being signed up for and abandoned? |
| Refusals | Are people being stopped by a limit, and which limit? |
| Plans | Which plan are people on, and are they moving? |
Refusals are counted separately from failures on purpose. A refusal means the installation worked and said no, because of a rate limit, a spend ceiling or a blocked model. A failure means something broke. Mixing the two makes a healthy installation look broken and a misconfigured one look busy.
Live activity
Insight → Pulse is the last minute: turns being served, requests in flight, and events as they arrive. It is
the screen to watch while changing something that affects everybody, because it answers "did that break
anything" in seconds rather than after a page refresh.
What to watch right after a change
Open Pulse, make the change in another tab, and watch for a minute. A misconfigured model or a tightened rate limit shows up here as either silence or a wall of refusals, and both are recognisable within seconds.
The three numbers worth knowing by name
Credits granted, spent, and in flight. In flight is the part that surprises people: a turn that is running has credits reserved against it, and those credits are settled when it finishes. An installation that looks like it spent more than it did is usually looking at reserved credits in the middle of a busy minute.
Turns per minute against the plan. Platform → Rate limits & abuse decides the ceiling per plan, and this
is where you see how close to it people actually run. A ceiling nobody reaches is decoration; one everybody
reaches is a support queue.
Refusals by reason. A refusal rate of a few percent during a launch is normal, because a busy installation has people hitting windows. A refusal rate that climbs while usage is flat is a setting, not a crowd.
Where the data comes from
Everything on these three screens is computed from the deployment's own records: the turns, the credit movements and the settings changes. There is no analytics service of anybody else's in the path, no script from another company on your users' pages, and nothing to accept a cookie banner for.
A dashboard is not a backup
Insight answers what is happening. It does not answer what happened a month
ago in detail, and it is not a substitute for System → Backups & exports
when you need to move or keep data. Take a snapshot before anything that
changes a lot at once.