What are evals?
Trying a prompt once and liking the answer tells you almost nothing. The model is partly random, real users ask hundreds of kinds of question, and a change that helps one kind often hurts another.
An eval is the answer to that. You collect a set of realistic questions, write down what a good answer must contain for each, and run all of them every time something changes. The result is a score you can compare before and after.
The best test questions come from real failures. Each time the product gets something wrong, that case joins the set, so the same mistake cannot return unnoticed.
Why do evals matter?
A change to a prompt can make one answer better and ten others worse, and nobody notices until customers do. Because a model's behaviour cannot be read from its prompt, the only way to know whether a change helped is to measure. Evals are the test suite of an AI product, and teams without them are guessing.
How do evals work?
- Collect real questions or tasks, each with what a good result looks like.
- Run the product on all of them.
- Score each result: by exact comparison, by a check in code, or by a judge model with a checklist.
- Compare the score with the previous version.
- Read the failures. They say what to fix. Add every new failure found in real use to the set.
When should you use evals?
Use it when
- Before any change to a prompt, model, tool or search setting
- Choosing between two approaches
- Checking that a fix did not break something else
Think twice when
- You are still exploring whether an idea works at all. Try it by hand first
Where do evals sit in a real system?
Evals run outside the live product, usually automatically whenever a change is proposed. The test set is kept alongside the code and grows over time.
What goes wrong with evals?
- Testing with five examples and calling it measured.
- Reading only the score and never the failures.
- Running each case once, when results vary from run to run.
What does evals look like in practice?
One sentence is added to an assistant's instructions and the first three answers look better. Run on 200 saved questions, the score falls from 91% to 86%. The new sentence made replies friendlier and made the assistant stop citing its sources.