Skip to main content
Create an experiment when you want to compare one or more candidates against the same dataset.

​
How do I create an experiment?

Open Experiments. Click Add Experiment. Enter a name and optional description. Choose a dataset if you already have one. Click Add Experiment. You can also create the experiment first and attach or change the dataset before running.

​
What should I add before running?

Before you start a run, make sure the experiment has:
  • a dataset version
  • at least one active candidate
  • input columns that match the prompt or agent inputs you expect to use
For prompts, make sure the dataset input column names match the prompt variables where possible. For agents, use columns that make the row intent clear, such as question, task, context, or customer_message.

​
How should I name experiments?

Use names that explain the comparison. Good examples:
  • Support FAQ prompt model comparison
  • Invoice extraction agent regression
  • Refund policy prompt v3 vs deployed
Avoid generic names such as Test or Experiment 1.

​
Editing an experiment

Open the experiment. Use the dropdown in the breadcrumbs to edit the experiment name or description. Use Archive from the same menu when you no longer need the experiment in the main list. Archived experiments are hidden from the default experiment list.

​
Reproducibility

Experiments are designed to be repeatable. Dataset versions are immutable. Candidates are captured as snapshots. This means a past run still points to the dataset rows and candidate configuration used at the time. If you change a prompt or agent later, add it as a new candidate or start a new experiment so the comparison remains clear. See also: Datasets, Add candidates, and Run an experiment