Evaluate a prompt against a dataset
Build a dataset from real user feedback or a CSV/JSON test set, run an experiment that sweeps prompt versions and models, and read the run report.
Improve a prompt from feedback
Turn replies your team disagreed with into candidate rewrites of your production prompt, score them against what is live, and promote the one that earns it.
Evaluate with conversation history
Carry a session's prior turns into a dataset example so a run, the judge, and the optimizer see the same conversation the model actually answered in.
Score live traffic
Create a standing rule that judges matching production gateway calls automatically, with no dataset and no manual run.
Evaluate a prompt that reads an image
Store a prompt with an image part, fill the image from a dataset example, run it on several vision models, and read the scores a judge gives after looking at the same image.
Write a custom judge prompt
Grade experiment runs and evaluation rules with your own rubric, including examples that have no criteria, and see exactly what the judge model receives.