عرض عملي: تشغيل تقييم يكشف تراجعًا في الـ prompt

خطوة بخطوة

1. Run the 50-question set on the current prompt: baseline recorded.
2. Open the report: recall, groundedness, and refusal correctness per category.
3. Planned mistake: shorten the prompt to save tokens (dropping the "answer only from sources" rule).
4. Run the pipeline: groundedness falls below the threshold; the build fails and lists the failing questions.
5. Restore the rule, keep the shorter wording elsewhere, and rerun: pass.
6. Open the dashboard: cost per conversation fell slightly, which would have hidden the regression.