Evals: What to test

For teams building an AI product who aren’t sure what their evals should check.

Overview

Start by reading A Field Guide to Rapidly Improving AI Products for an overview of the process you can use to figure out what to test.

Then watch Intro to Error Analysis With Just Spreadsheets for a walkthrough of error analysis, or read the short FAQ, How is error analysis performed?. Evals skills can help you do this work with a coding agent.

Watch Shreya’s How to Automate AI Evals (Correctly) to see the Error Discovery skill in action.

Common questions

Sampling and generating data

• How can I sample production traces efficiently?
• How many examples do I need?
• How do I find failures beyond user feedback?
• How should I generate synthetic data?
• When is synthetic data unreliable?

Reviewing and labeling outputs

• What if traces contain sensitive data?
• What if users ask very different kinds of questions?
• Do I need reference answers or a rubric first?
• How many people should annotate outputs?
• Should I record problems that aren’t the model’s fault?
• How should PMs and engineers collaborate?
• Can I help without domain expertise?

Deciding what to test

• Should I build automated evaluators for every failure?
• Should I write evals before building a feature?

← See all the guides