Handling tool use responses

Tool Use with LLMs (Claude)

🛠️ Configuring Tools in the Request

• Tool Inclusion: When making a request to Claude, include the JSON schema specification that describes how to call your tool.
• tools Argument: Add a tools keyword argument to the chat function, which accepts a list of your tool schemas.
• tool_config (Advanced Control): This nested dictionary controls how the model decides to use the provided tools.
• auto (Default): The model delegates the choice of whether or not to use a tool to itself.
• any (String): Forces Claude to use a tool, and it decides which one.
• Specific Tool Name (Nested Dictionary): Forces Claude to call a specific tool by providing its name.

🤖 Understanding the Model Response (Tool Use)

When the model decides to use a tool, the response structure changes significantly:

• stop_reason: This key indicates why the model stopped generating output.
• If the model decides to use a tool, the stop_reason will be tool_use.
• Assistant Message Structure: The assistant message's content list will contain multiple parts, not just text.
• Text Part: Contains any helpful introductory text from the model (e. g. , "I can check the current date time for you. ").
• Tool Use Part: A specific dictionary indicating the model's intent to run a tool.
• name: The exact name of the tool from your JSON schema that the model wants to run.
• input: A dictionary containing the arguments that the model wants you to pass into your tool function.

🔄 Necessary Code Refactoring

• Handling Multi-Part Messages: Since the assistant message now contains multiple parts (Text + Tool Use), the existing message handling functions must be updated.
• chat Function Update: The function must now return both the extracted text and the full list of message parts.
• Message Management Functions (add_user_message, add_assistant_message): These functions were refactored to be flexible:
• They must accept either a simple string (for basic text) or a list of parts (for complex tool-use responses).
• The logic was updated to correctly wrap a single string into a single text part when necessary.

Takeaways

• Tools are configured by including a JSON schema specification within the tools argument of the request.
• When a tool is used, the model response features a stop_reason of tool_use and a multi-part assistant message.
• The tool use part of the message contains the exact tool name and a dictionary of required input arguments.
• Message handling functions must be updated to accept and process messages that are either a simple string or a complex list of parts.