Azure AI Foundry: Retrieving files of Code Interpreter tool (Responses API)

This repo demonstrates how to use Azure OpenAI's Responses API with the Code Interpreter tool. The Code Interpreter enables your selected GPT model to build and test Python code within a secure sandbox container, generating outputs such as data visualisations and analysis results.

Using REST API calls (via the requests library), we can retrieve files generated by the Code Interpreter's Python run, including images, charts and other outputs.

Note

For the latest implementation details of Responses API in Azure, refer to the Azure AI Foundry documentation page.

📑 Table of Contents:

• Part 1: Configuring Solution Environment
• Part 2: Calling Code Interpreter Tool
• Part 3: Retrieving Output Files from Container

Part 1: Configuring Solution Environment

To use this notebook, you'll need to set up your Azure OpenAI environment and install the required Python packages.

1.1 Azure OpenAI Service Setup

Ensure you have an Azure OpenAI Service resource with a model deployment that supports the Code Interpreter tool calling (e.g., GPT-4.1).

1.2 Authentication

This demo uses Microsoft Entra ID authentication via DefaultAzureCredential from the azure.identity package.

To enable authentication, ensure your environment is properly configured by:

• Logging in via az login (Azure CLI), or
• Setting relevant environment variables for service principals, or
• Using managed identity in Azure environment.

Define a token provider using the get_bearer_token_provider() function:

token_provider = get_bearer_token_provider(
    DefaultAzureCredential(),
    "https://cognitiveservices.azure.com/.default"
)

1.3 Environment Variables

Configure the following environment variables for your Azure OpenAI deployment:

Environment Variable	Description
AOAI_API_BASE	Your Azure OpenAI endpoint URL (e.g., https://<YOUR_AOAI_RESOURCE>.openai.azure.com).
AOAI_DEPLOYMENT	The name of your model deployment (e.g., gpt-4.1).
AOAI_API_VERSION	The API version (e.g., 2025-04-01-preview).

1.4 Installation of Required Python Packages

Install the necessary packages:

pip install openai azure-identity requests Pillow ipython pathlib

Part 2: Calling Code Interpreter tool

2.1 Initialisation of Azure OpenAI client

Initialise the Azure OpenAI with your environment variables and Entra ID token provider:

client = AzureOpenAI(
    azure_endpoint = AOAI_API_BASE,
    azure_ad_token_provider = token_provider,
    api_version = AOAI_API_VERSION,
)

2.2 Responses API with Code Interpreter tool

Include the Code Interpreter as a tool in your Responses API call, along with instructions (system prompt) directing the model to use the appropriate tool:

response = client.responses.create(
    model = AOAI_DEPLOYMENT,
    tools = [
        {
            "type": "code_interpreter",
            "container": {"type": "auto"}
        }
    ],
    instructions = "You are a helpful data analyst. You should use Python tool to perform required calculations.",
    input = INPUT_TEXT
)

Part 3: Retrieving output files from container

3.1 Details of Generated Files

When the Code Interpreter generates output files, their details (Container ID, File ID and filename) can be extracted from the "annotations" section of the Response API output:

{
    "container_id": "cntr_689df3cb69648190b73217f54eeb713806df641648828556",
    "file_id": "cfile_689df41900fc81909d22ec83e7dafbe1",
    "filename": "cfile_689df41900fc81909d22ec83e7dafbe1.png"
}

Warning

Details shown above are examples. Your actual container ID, file IDs and file names may be different for each execution.

3.2 REST API for Container Files' Content

Retrieve the content of generated files using the REST API endpoint that references the temporary container ID and each file ID:

{AOAI_API_BASE}/openai/v1/containers/{container_id}/files/{file_id}/content

3.3 File Download and Display

After retrieving and saving the file content bytes locally, you can display the outputs in your Jupyter notebook. For example, a data visualisation might look like this: