Reference
Concepts and glossary
Understand the resources used across the dashboard and integrations.
| Term | Meaning |
|---|---|
| Organization | The workspace that owns shared resources, membership, providers, subscription, and credits |
| Prompt | A stable resource ID with messages, settings, versions, labels, a dataset, and evaluators |
| Draft | Your personal editable prompt configuration, protected by a revision |
| Saved version | An immutable snapshot of prompt configuration with a version ID and number |
| Label | A movable name, such as production, pointing to a saved version |
| Dataset | The prompt's shared, ordered collection of test cases |
| Case | An input plus an optional reference output and independent field formats |
| Reference / expected output | The desired answer used by matching or review; named expectedOutput in the API |
| Evaluator | A configured rule scoped to all or selected cases of a prompt |
| Evaluation / run | Background work that records a version's outputs and scoring against captured inputs |
| Baseline | The initial evaluated version, or the production result used as a comparison reference |
| Candidate | The saved version being compared with production |
| Coverage | Which dataset cases and dataset state an evaluation represents |
| Model route | A model combined with the provider route used to serve it |
| Judge | A model that evaluates another model's output using specified criteria |
| Revision | A concurrency value used to prevent overwriting a newer edit |
| Retrieval key | A prompt-scoped credential that reads saved prompt configuration |
| Management token | An OAuth credential for authorized account and organization operations |
IDs, numbers, and revisions
A version ID identifies the saved object used by management operations. A version number such as 1 is the selector used by application retrieval. A draft, dataset, or label revision protects mutable state; it is not a saved prompt version number.
Configuration and execution
Retrieval returns a saved prompt without executing a model or rendering input variables. Your application assembles input messages and calls the provider. Evaluations execute model calls against a captured dataset and record the outcomes. These are separate operations with different credentials and costs.
For a practical sequence connecting the concepts, follow your first evaluation and production publication.