Level 2: Defined¶
Last updated: 2026-07-27
List of references
At Level 2, the team moves from individual experiments to a defined engineering approach. Prompts, grounding configuration, application code, evaluation data, and deployment definitions are versioned and reviewed together.
Define the delivery path¶
| Capability | Defined practice |
|---|---|
| Prompt lifecycle | Store system prompts, templates, and configuration in Git with pull-request review. |
| Evaluation | Run a repeatable evaluation set before promotion and retain results with the release. |
| Safety | Include content safety and scenario-specific policy tests in the release gate. |
| Deployment | Use a repeatable pipeline for dev, test, and production environments. |
| Secrets | Use managed identities and Azure Key Vault rather than application configuration files. |
Source: Develop flows in Azure AI Foundry.
Evaluation gate¶
Establish an evaluation dataset containing representative requests, adversarial inputs, known failure cases, and desired answers or judging criteria. The gate should compare the candidate against the current production prompt and model, not merely a generic minimum score.
Business example: A support assistant can only be promoted when grounded answer quality remains above the approved threshold, unsafe-answer findings do not increase, and the release stays within the response-time budget.
Integration and deployment¶
Integrate the application through a controlled service boundary such as Azure Functions, Azure Container Apps, or Azure App Service. Continuous integration should run linting, unit tests, prompt tests, and evaluation; continuous delivery should deploy to a non-production environment before an approved promotion.
Move to Level 3: Managed when the application has a repeatable delivery process and needs production operating controls.