Product guide · English
Evaluate AI prompts and models
Use the playground to learn the evaluation flow, then configure real targets, test cases and graders before running a saved evaluation.
Steps
- Try prompt variations in the playground with synthetic inputs. This is a local demo: no real model is connected and results are not saved as production evaluation evidence.
- Create an evaluation and define the prompts and test cases. Add a model connection or an HTTP target under Targets, and confirm what data that target will receive.
- Choose rule checks, an LLM judge or semantic-similarity grading as appropriate. Review the grader settings and expected outputs before starting a run.
- Inspect the run’s results matrix, failures, pass rate and cost. Compare like-for-like inputs and configurations; use schedules only after checking the intended recurring scope.
Read the result
Pass rates summarize performance on the chosen cases and graders. Judge and similarity outputs are evidence to inspect, not ground truth. Keep prompt, target and grader changes explicit when comparing runs.
Limits and availability
Real runs depend on configured targets, access and usage limits; they can send inputs to the selected provider or endpoint. The playground does not establish real-model quality, latency or cost. Use approved test data and do not infer production reliability from a small sample.
This guide describes the implemented workflow. The features available in your workspace depend on account capabilities and configuration; advanced workflows are not all included in Free Quick Scan. Check the available controls and the evidence returned by each run.