Question

Your team deployed a chat application backed by a foundation model served through an Azure AI Foundry serverless API endpoint. The deployment currently targets the model by its base name without specifying a version. After the provider auto-upgraded the model, users report subtle changes in output formatting that broke a downstream parser. The compliance team requires that production behavior remain reproducible and only change through a controlled review. What should you do to prevent unexpected model behavior changes in production?

Accepted Answer

Pin the deployment to a specific model version and promote new versions only after validating them through your evaluation pipeline.

Answer

Enable provisioned throughput units on the endpoint so the model version is frozen automatically.

Answer

Increase the endpoint's request timeout so slower responses from newer model versions do not fail.

Answer

Switch the deployment from a serverless API endpoint to managed compute to disable automatic model updates.