In plain language
They are the test suite for an AI workflow, showing whether a prompt or model change made results better or worse.
Example
Before changing a classification prompt, it is run against 200 past tickets with known correct categories.
Related terms
Classification · Large Language Model (LLM) · Guardrails · Hallucination
Common misconception
A few successful manual tries are not an eval; results need to be measured across many representative cases.
Implementation guidance
Build a test set from real cases including hard ones, define pass criteria, re-run evals after every prompt or model change, and track results over time.