4.1.2. Deploying and Interacting with a Model in Foundry
💡 First Principle: Before you can call a model, it must be deployed — selected from the Foundry model catalog and provisioned as an endpoint in your project. Deployment turns "a model that exists" into "a model I can send requests to," with a deployment name you reference in code or in the playground.
In the portal, the flow is: browse the model catalog, deploy a model (giving it a deployment name), then open it in the playground to test prompts interactively without writing code. The playground is where you'd tune the system prompt and parameters like temperature before committing to an app. When you're ready for code, you connect with the project endpoint and your Azure credentials.
⚠️ Exam Trap: The model name and the deployment name are not always the same thing. You deploy (say) a GPT model and give that deployment your own name; your code references the deployment name. A "model not found" error often means the deployment name is wrong, not the model.
Reflection Question: Why does Foundry make you deploy a model before using it, rather than letting you call any catalog model directly by its name?
