In 30 seconds
- AI features are non-deterministic; unit tests alone aren't enough.
- Evals are curated examples plus scoring rules.
- Write evals first, then iterate prompts and models against them.
Why evals
Traditional tests expect exact outputs. AI features give varied outputs that can be good or bad in degrees. An eval is a set of representative inputs with a way to score the outputs: rules, reference answers, or human or AI grading.
Building an eval set
- Collect real or realistic examples, including hard cases.
- Define what good looks like for each.
- Choose scoring: exact match, rules, rubric, or graded review.
- Set a pass threshold before building.
Evals in agile delivery
Evals play the role of acceptance criteria and TDD tests for AI features. Run them on every change in the pipeline, add failing cases from production as you would escaped defects, and include them in the Definition of Done.
Myth vs reality
Myth · A few manual checks are enough.
Reality · AI behaviour varies; you need systematic evals.
Myth · Evals are a one-off setup.
Reality · They grow with every failure you find.
Myth · Higher eval scores always mean better.
Reality · Overfitting to the eval set is a real risk.
FAQ
How many examples do we need?
Start with dozens covering key cases; grow over time.
Can AI grade AI output?
Yes, with care and periodic human checks.
Who owns evals?
The team, with the Product Owner defining what good means.