This page describes how to use tool calling in Agent Builder. If you’re looking for an overview of tool calling, see Introduction to tool calling.

Agent Builder provides two ways to use tool calling:

Both blocks allow you to either connect to Knowledge Graphs or define the functions you want to provide to the model as JSON Schema objects. Then, when the model decides to use a tool, it either queries the Knowledge Graphs or sends the call to your blueprint, which executes the tool and sends the results back to the model.

The difference between these the Generate chat reply and Tool calling blocks is that the Generate chat reply block automatically adds the results from the tools to the response and adds it to the chat history of an ongoing conversation, while the Tool calling block allows you to call tools outside of a chatbot conversation.

​
Add tool calling to your blueprint

To add tool calling to your blueprint, you need to:

  1. Add the Tool calling block to your blueprint
  2. Write a prompt that explains what the tool calling block is trying to accomplish
  3. Define the tools you want to provide to the model
  4. Connect the tool calling block to the blocks that execute the tools
  5. Add a Return value block to the end of the tool calling block

​
Add the tool calling block

Add either the Tool calling block to your blueprint or the Generate chat reply block to your chatbot conversation.

​
Write a prompt

The prompt you write for the tool calling block in the Prompt field guides the model’s behavior and decision making. It should explain what the tool calling block is trying to accomplish and what tools are available to use.

​
Define your tools

Define the tools you want to provide to the Tool calling block, so that it knows what tools are available to use.

To add a new tool, click the Add tool+ button in the Tool calling block’s configuration menu.

In the Add tool modal that appears, select the Tool type from the dropdown, either:

  • Function: A function that the model can call
  • Knowledge Graph: A Knowledge Graph that the model can query

If you’re using a function, enter the function name and definition. See tool definition for more details about how to define the function.

If you are using the Knowledge Graph tool, you only need to select which Knowledge Graphs you want to use. You do not need to define the tool or connect it to the blocks that execute the tools as described below. The following steps are only for the Function tool type.

​
Tool definition

The tool definition for a Function tool follows the JSON Schema format and has the following structure:

ParameterTypeDescription
namestringThe name of the tool
descriptionstringA description of what the tool does and when the model should use it
parametersobjectAn object containing the tool’s input parameters
parameters.typestringThe type of the parameter, which is object for a JSON schema
parameters.propertiesobjectAn object containing the tool’s parameters in the form of a JSON schema. See below for more details.
parameters.requiredarrayAn array of the tool’s required parameters

​
Example tool definition

Below is an example of a tool definition that gets the details of a package’s shipping status. It takes a tracking number and a carrier name as input parameters and returns the shipping status of the package.

{
  "name": "check_shipping_status",
  "description": "Gets real-time shipping and tracking information for a package",
  "parameters": {
    "type": "object",
    "properties": {
      "tracking_number": {
        "type": "string",
        "description": "The shipping carrier's tracking number"
      },
      "carrier": {
        "type": "string",
        "enum": ["fedex", "ups", "usps", "dhl"],
        "description": "Shipping carrier name"
      }
    },
    "required": ["tracking_number", "carrier"]
  }
}

​
Connect the tool calling connectors to the blocks that execute the tools

When you define a Function tool in the block’s configuration, it adds a connection point (the purple circle) to the block. This allows you to connect it to another block or chain of blocks that perform the tool’s work.

The example above shows the following setup for a tool that, among other functions, can check the status of a package:

1

Create an HTTP Request block that calls a shipping status API.

2

Add a Return value block to the end of the HTTP Request block to send the result back to the model.

3

Provide a tool definition to the Tool calling block that indicates there’s a tool that can check the status of a package and it needs the tracking number and carrier name as input parameters.

4

Connect the HTTP Request block to the Tool calling block at the check_shipping_status connection point.

When the model decides to use the tool, it sends the tool call to the HTTP Request block, which executes the tool. The HTTP Request block runs and passes the result to the Return value block, which sends the results back to the model

Every Function tool call needs a Return value block at the end that sends the results back to the model.

When the Tool calling block completes all tool calls and integrates the results into the response, it moves to the next block in the blueprint that’s connected to the Success transition. The @{result} variable is available in the next block with the tool calling block’s final output.

​
View tool calls logs

When a Tool Calling block executes, it provides a list of thoughts and actions that it took to complete the task. You can observe these in the logs of the blueprint.

Expand the Logs section at the bottom of the blueprint. Find the Tool calling log and click the Trace button to see the full trace of the tool calls.

The trace shows the series of thoughts, actions, and tool calls that the model made to complete the task.