Anvia

Production checklist ​

Treat the Azure client as infrastructure and each deployment as a tested application dependency.

Configuration ​

  • Keep keys, token providers, and clients in server-only modules.
  • Prefer Microsoft Entra tokens (azureADTokenProvider) over long-lived keys where your platform supports managed identity.
  • Select the token scope for the endpoint: https://cognitiveservices.azure.com/.default for resources and https://ai.azure.com/.default for Foundry projects.
  • Use Azure-specific variable names (AZURE_OPENAI_*) so OpenAI and Azure configuration cannot be mixed up.
  • Keep deployment names in trusted configuration and allow-list the ones users may select.
  • Declare contextLimits and controls for custom deployment names.

Capability tests ​

  • Smoke test the exact endpoint, API (responses or chat), and deployment.
  • Exercise agent.generate() and agent.stream() if the product uses both, including a tool call with complete streamed arguments.
  • Validate output-schema behavior before relying on parsed product data.
  • Test embeddings and media with representative input against their own deployments.
  • Treat model listing as endpoint inventory, not proof of a working deployment.

Reliability ​

  • The managed client disables SDK retries. Bound request time, retries, and concurrency in your Agent or application policy.
  • Retry only transient failures and use idempotency around downstream writes.
  • Avoid silently switching provider, API, or deployment after a failure.
  • Record azure-openai and the deployment in traces so regressions and spend can be attributed.

Migrating from OpenAIClient ​

If you previously pointed OpenAIClient at Azure with baseUrl, move to AzureOpenAIClient and keep the same model-handle calls. The existing custom-endpoint path still works but does not include Azure Responses event normalization. See Migrate from @anvia/openai.

Security and data ​

  • Azure data handling, region, and retention are set by your Azure resource and deployment, not by Anvia.
  • Do not log prompts, tool arguments, or credentials by default.
  • Keep authorization in tools and application code, not in the model's instructions.

Built for Anvia.