Evaluation Center
The Evaluation Center uses reusable cases, criteria, and facts to assess position responses. Session scanning also finds semantic issues across real conversations. Prepare evaluation data, run evaluations, and inspect reports. Use diagnosis for individual issues or create a session scan for issues across conversations.
Recommended workflow
- Prepare Cases, Criteria, and Facts
Create cases, test sets, evaluation criteria, and factual baselines.
- Configure and Run Evaluations
Select targets, test sets, and criteria, then run a batch evaluation.
- View Evaluation Reports
Compare batches and identify failed cases and quality changes.
- Diagnose Answer Issues
Inspect timing, tool calls, and model interactions to locate answer issues.
- Session Scanning
Select real historical conversations and generate semantic improvement suggestions in bulk.
Understand this capability
An evaluation has four parts: two for preparation and two for execution and review.
| Component | Purpose |
|---|---|
| Cases / Test sets | What to evaluate: reusable groups of questions, optionally with reference answers |
| Evaluation criteria | How to evaluate: scoring definitions and pass thresholds for each metric |
| Evaluation tasks | Which configuration to run: test set, execution position, execution model, and evaluation model |
| Evaluation reports | How to interpret results: batch overviews, case details, fact verification, and batch comparisons |
The full flow is: import cases and form a test set → configure a task → run evaluation → review results → improve related configuration → rerun the same cases.