Your First Eval
Dataset in, scores out — measure a template instead of eyeballing it.
Why evals first
Before publishing a template or changing its prompt, run it against known cases. Evals turn "feels better" into numbers.
Steps
- Seed a dataset — easiest from the template itself:
POST /api/v1/templates/{id}/create-dataset(evaluation.py:661), or build one manually viaPOST /datasets(evaluation.py:80). - Add test cases with expected outcomes — bulk import helps
(
evaluation.py:188). - Launch a run (
POST /runs,evaluation.py:258). Cases execute like missions; costs apply. - Read scores & costs on the run dashboard, then A/B two prompt
variants with
POST /compare(evaluation.py:285).
Next: publish confidently with Publish a Template.
Last updated 2026-08-25 (git-derived)