What are tools?

Microsoft Foundry Models includes models that are capable of using tools to find information or perform tasks. You can use tool support in models by specifying which tools you want the model to use in prompts submitted through the OpenAI Responses API.

When you develop a generative AI application using Microsoft Foundry, you can search Foundry Models for a model with tool calling capabilities and deploy it. Then, you can develop client applications that use the OpenAI Responses API to submit prompts to the deployed model, specifying the tools that the model can use.

Some of the commonly used tools available in the Responses API, include:

• code_interpreter: A Python environment in which the model can generate and run code.
• web_search: A tool that enables the model to find general information on the Internet, which allows it to base responses on more current data than it was trained on.
• file_search: A tool that enables the model to search specific files that you upload to a dedicated vector search index - enabling it to ground responses in specific knowledge.
• function: A tool that enables the model to call custom functions in your application code.

We'll explore these tools in this module.

Specifying tools in the Responses API

You can specify one or more tools in a call to the responses.create() method when generating a response from a model. The following Python pseudocode example indicates where the list of callable tools is specified:

from openai import OpenAI

client = OpenAI(
    base_url={openai_endpoint},
    api_key={auth_key_or_token}
)

response = client.responses.create(
    model={model_deployment},
    instructions="You are a helpful AI assistant.",
    input="Find me some information about vintage computers.",
    # Specify available tools as a JSON list
    tools=[
        { 
            # A tool definition
            "type": "{tool_type}",
            "{tool-specific-setting}": "{value}",
                ...
        },
        { 
            # Another tool definition
            "type": "{another_tool_type}",
            "{tool-specific-setting}": "{value}",
                ...
        }
    ]
)
print(response.output_text)