Skip to main content
Drift monitors watch a metric or output feature over time and alert when it diverges from a baseline distribution.

​
How it works

  • You pick a target — an output feature (e.g. total_tokens, a score, a latency) — and a window (e.g. hourly).
  • RunAgain computes a drift statistic per window against the baseline. The default method is PSI (Population Stability Index).
  • When the statistic crosses your threshold, the window is flagged warning or drift.

​
Creating a monitor

In a project, open Drift → New monitor and set:
  • Target — the feature/metric to watch (e.g. total_tokens).
  • Method — psi (default).
  • Window — granularity and size (e.g. 1 hour).
  • Threshold — the value + operator that trips an alert (e.g. >= 0.25).

​
Reading the series

Each monitor shows a time series of the drift statistic with ok / warning / drift status per window, plus sample sizes so you can tell signal from noise.

See it in the app

Drift monitors live under each project’s Drift tab.