Ground truth
The known correct answers or labels that a model’s output is compared against in an evaluation.
For classification it is the right label; for RAG it is the reference answer and the documents that contain it. Metrics such as recall or context recall cannot be computed without it. It usually comes from people who know the domain, which makes it the expensive part of an evaluation, and it can itself be wrong or out of date. A golden dataset is ground truth packaged with inputs for repeated runs.
Related terms
- Golden dataset: A fixed set of real inputs with expected answers or checks, used to test an AI feature after every prompt or model change.
- Evals: A repeatable test set that scores an AI feature, so changes can be measured instead of guessed.
- Context recall: A RAG retrieval metric: how much of the information needed for the correct answer was actually retrieved.
- Human evaluation: People, ideally domain experts, grading AI outputs against a written rubric; the reference that automated metrics are checked against.
