LLM outputs vary, and small changes have large effects. That makes an LLM feature hard to change safely. Evals are how you measure whether the system works, and they are what lets you change anything with confidence.
Why LLM systems are hard to test#
Three things make behavior shift under you:
- The same prompt varies. Run it twice and you may get slightly different answers, as the temperature note explains.
- Small prompt changes break things. Change one word in the system prompt and something that worked before may stop working.
- Model upgrades shift behavior. A new model version can behave differently in ways you did not expect.
Without a way to measure quality, you cannot tell whether a change made things better or worse. Every prompt tweak, model upgrade or chunking change is a guess, and regressions reach production unnoticed.
What evals give you#
An eval is a systematic way to measure whether your system is working. You run a fixed set of test cases, check the outputs, and get a result you can compare across versions.
"I tested it manually and it looked good" does not scale. Manual testing covers the happy path. Evals also cover edge cases, refusals, and the cases users will actually hit that you did not think of.
| Manual testing | Evals | |
|---|---|---|
| Coverage | Happy path | Happy path, edge cases, refusals, adversarial |
| Repeatable | ✗ | ✓ |
| Compares versions | ✗ | ✓ |
| Catches regressions | Only by luck | Every run |
The rest of this category builds an eval system piece by piece: the golden set of test cases, simple metrics in code, an LLM judge for fuzzy quality, regression runs on every change, and signals in production.
Short line#
LLM outputs vary and small changes break things, so without evals every change is a guess. Evals measure quality on a fixed set of cases, cover more than manual testing, and make changes safe.
Key takeaways#
- The same prompt can give different answers, and small prompt or model changes shift behavior.
- Without measurement, you cannot tell whether a change helped or hurt.
- Evals run fixed test cases so you can compare versions.
- Manual testing covers the happy path. Evals cover edge cases, refusals and unexpected cases too.