The outcome
Operate a generative AI flow that has evidence-based quality thresholds and documented responses to unsafe or unsupported content.
Step by step
A workflow you can repeat.
- 01
Create a versioned dataset covering normal tasks, prompt injection, sensitive data, harmful content, protected material, and unsupported-answer cases.
- 02
Define task, groundedness, safety, latency, and cost metrics with pass thresholds and a human review rubric for ambiguous results.
- 03
Configure system rules, content filters or Content Safety checks, Prompt Shields, network boundaries, and output validation for the use case.
- 04
Run evaluations on the complete flow, inspect false positives and false negatives, and test tool calls and retrieved content as untrusted inputs.
- 05
Version prompts, filters, models, and datasets together; monitor production signals, audit changes, and block promotion when regression thresholds fail.
Working standard
What good use looks like.
- Evaluate the complete application flow.
- Measure false positives and false negatives.
- Version guardrails with prompts and models.
Official references