Edit

Share via


Quickstart: Get answers in the chat playground

Note

This document refers to the Microsoft Foundry (classic) portal.

🔍 View the Microsoft Foundry (new) documentation to learn about the new portal.

Learn how to use the chat playground in Microsoft Foundry to explore AI model capabilities interactively. This quickstart focuses on the web-based UI experience; to build applications programmatically, see Build a custom chat app using the SDK.

Deploy (or reuse) a chat model and send prompts to receive AI-generated responses.

In this quickstart, you learn how to:

  • Configure a system message to guide model behavior.
  • Send a user question and receive a response.
  • Interpret model responses and recognize limitations.
  • Add safety system messages to ensure responsible AI use.

For this quickstart, you can use either a hub-based project or a Foundry project. For more information about the differences between these two project types, see Project types.

Prerequisites

An Azure account with an active subscription. If you don't have one, create a free Azure account, which includes a free trial subscription.

Deploy a model

In the portal, you can explore a rich catalog of cutting-edge models from many different providers. For this tutorial, search and then select the gpt-4o model.

  1. Sign in to Microsoft Foundry. Make sure the New Foundry toggle is off. These steps refer to Foundry (classic).

  2. If you're in a project, select Microsoft Foundry in the upper-left breadcrumb to leave the project. You'll create a new one in a moment.

  3. From the landing page or Model catalog, select gpt-4o (or gpt-4o-mini).

    Screenshot shows how to start with a model in Foundry portal.

  4. Select Use this model. When prompted, enter a new project name and select Create.

  5. Review the deployment name and select Create.

  6. Then select Connect and deploy after selecting a deployment type.

  7. Select Open in playground from the deployment page after it's deployed.

  8. You land in the Chat playground with the model pre-deployed and ready to use.

If you're building an agent, you can instead start with Create an agent. The steps are similar, but in a different order. Once the project is created, you arrive at the Agent playground instead of the Chat playground.

You need the following permissions and setup:

Use the chat playground

Use the Foundry playground to interact with deployed chat models and test prompts in real time.

To get answers from your deployed model in the chat playground:

  1. In the System message text box, provide a prompt to guide the assistant. For example, for a customer support scenario, use: "You're a helpful customer support agent. Answer questions about product features, pricing, and troubleshooting. If you don't know the answer, offer to escalate to a specialist." You can tailor the prompt for your specific use case.

  2. Optionally, add a safety system message by selecting the Add section button, and then Safety system messages. Choose from the prebuilt messages, and then edit them to your needs.

  3. Select Apply changes to save your changes. When prompted to see if you want to update the system message, select Continue.

  4. In the chat session pane, enter the following question: "How much do the TrailWalker hiking shoes cost?"

  5. Select the right arrow icon to send.

    Screenshot of the first chat question without grounding data.

  6. The assistant either replies that it doesn't know the answer or provides a generic response, such as noting price variability. This is expected because the model doesn't have access to current product data.

Understanding the response: The model generated text based on its training data and system message, but without grounding data (like a product catalog), it can't provide accurate domain-specific answers. This limitation is normal and expected in this scenario.

Next, add your data so the model can answer domain-specific questions. Try the enterprise chat web app tutorial.

Troubleshooting

Issue Action
No deployed models listed Deploy a model from the model catalog first.
Repeated generic answers Refine system message or add domain data.
Safety message overrides tone Adjust or remove conflicting safety sections.
Slow first response Allow for cold start; subsequent prompts are faster.

Next step