Quick recall
Before choosing an eval metric, what should you define first?
Say your answer first, then reveal the explanation.
Define the user task and what a good result must do. Then choose checks that measure those expectations.
AI evaluations · Lesson 1
Estimated time: about 10 minutes
An eval does not reveal whether an AI is “correct” in every possible situation. It checks whether the system meets clear expectations on a chosen set of cases. The choice of cases and expectations is part of the work.
Suppose you build an assistant that answers questions from a rulebook. “Is the AI good?” is too broad to check. Instead, ask: “When a user asks about a rule, does the assistant give the right ruling and point to the right source?” That question can be turned into examples and observable checks.
| Question in the eval | What a case might contain | What a pass could mean |
|---|---|---|
| Can it answer a supported question? | Question plus known source and key facts | Answer includes the required rule and cites its source |
| Does it avoid guessing? | A question the reference material cannot answer | It says it cannot determine the answer from the material |
| Does retrieval find the evidence? | Question plus the documents needed | At least one required document appears in the retrieved results |
One eval can have many cases. A grader is the rule or reviewer that decides whether a case passed. For an exact requirement, such as “the response must include a citation,” code can check it. For meaning, such as “the answer follows the cited rule,” a written rubric and human review may be needed. A model can also apply a rubric, but its judgments should be checked against human ratings.
Imagine 20 questions: 18 supported questions and 2 that should receive “I don’t know.” The assistant answers all 18 supported questions correctly, but also invents answers to both unsupported questions. A single score of 90% may look good, yet it hides a serious guessing problem.
So report the relevant kinds of success separately. For this example, show supported-answer accuracy and unsupported-question abstention accuracy. Then read the two failing answers. The overall score tells you how often cases passed; the examples help you understand why.
1. Your assistant must return valid JSON with a field named total. What is a strong first check?
total exists. This checks the exact format requirement without asking a language model to judge it.2. The score is 95%, but users report made-up answers to questions outside the source material. What should you inspect next?
The example set is called an evaluation dataset. It is a measuring instrument, so it can be incomplete or biased. A high score means “this system did well on these cases under these checks.” It does not guarantee correctness on future requests.
Start with OpenAI’s Evaluation best practices, which explains how to define success, select cases, choose graders, and improve evaluations over time. For the platform-specific workflow, see Working with evals.
Keep the steps handy in the AI eval design reference. Ask me any follow-up question; we can work through an example from a system you care about.
Answer before you reveal the explanation.
Quick recall
Say your answer first, then reveal the explanation.
Define the user task and what a good result must do. Then choose checks that measure those expectations.
Quick recall
Say your answer first, then reveal the explanation.
Measure accuracy on answerable cases and correct abstention on unsupported cases separately, then inspect the unsupported failures.
Quick recall
Say your answer first, then reveal the explanation.
It establishes that every case in that dataset passed the chosen checks in that run. It does not guarantee correctness on other requests.