'No Crash' and 'Correct Answer' Are Different Things

The same discipline that makes a unit test suite valuable — a fixed, repeatable set of checks run every time code changes — applies directly to AI feature quality. A small eval set of realistic questions with defined expected facts catches quality regressions a one-time manual spot check would miss.

Judging an AI feature's output requires a separate, deliberate comparison against a defined expectation — not just re-reading the answer and deciding it 'looks fine.' Keeping the grading step explicit and separate from generation is what makes the evaluation trustworthy.

'It Works' Isn't a Quality Bar for AI Features. A traditional feature either throws an error or it doesn't — pass/fail is usually clear. An AI feature can run successfully every time and still give a wrong, unhelpful, or subtly misleading answer. Evaluating quality needs its own deliberate process, closing this module's loop.

Build a Small Eval Set. The core technique: a small, fixed set of real questions with known-good answers, checked whenever the feature changes. Draft a starting eval set for the FAQ search feature.

Why is a small, fixed eval set more valuable than occasionally trying a few questions manually and eyeballing the answers?

• →A fixed eval set gives a repeatable, comparable check you can re-run every time the prompt, model, or retrieval logic changes — catching regressions that ad hoc spot checks would miss or forget to recheck.
• →Because it needs to contain thousands of questions to be useful, which manual checking could never achieve.

Run the Eval and Judge the Real Output. Take one of the drafted questions, get the model's real answer, and judge it against the expected key facts — this is the actual evaluation step, done by hand at small scale before any automation.

This Loop Repeats Every Time You Change the Feature. Every time the prompt, retrieval logic, or underlying model changes, re-running this small eval set tells you whether quality improved or regressed — turning 'I think it seems better' into an actual, checkable comparison. This closes the AI-Native Products module, and the second build-along module of this course.