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, Why is "error analysis" so important in AI evals, and how is it 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 efficiently sample production traces for review?
• How many examples do I need for an eval?
• How do I surface problematic traces for review beyond user feedback?
• What is the best approach for generating synthetic data?
• Are there scenarios where synthetic data may not be reliable?

Reviewing and labeling outputs

• How can I do evals when traces contain sensitive data?
• How do I approach evaluation when my system handles diverse user queries?
• Do I need a reference answer or rubric before annotating data?
• How many people should annotate my LLM outputs?
• Should I record problems that aren’t the model’s fault?
• Should product managers and engineers collaborate on error analysis? How?
• Can I help with evals if I’m not a domain expert?

Deciding what to test

• What model or LLM should I use to build automated evals?
• How should I evaluate a coding agent?
• Should I build automated evaluators for every failure mode I find?
• Should I practice eval-driven development?

← See all the guides