Multi-Turn conversations with tools

Key Concepts for Tool-Using Conversations

• The Problem with Simple Queries: When a query does not require tool use (e. g. , "what is one plus one"), the system must correctly identify the end of the conversation to avoid erroneously adding empty tool-use messages.
• The Role of stop_reason: To manage long-running conversations involving tools, the system must inspect the stop_reason returned in the API response. This reason dictates whether the conversation needs further processing or tool execution.
• Refactoring the chat Function:
• The chat function must be updated to return a structured dictionary, not just text.
• The returned dictionary must include:
• parts (the list of message parts).
• stop_reason (the reason the model stopped generating).
• text (a consolidated string created by joining all text parts from the message).
• Implementing the Conversation Loop (run_conversation):
• A multi-turn conversation requires a loop (e. g. , while True) to handle the back-and-forth between the model and the tools.
• Loop Logic:
• Call the chat function with the current message history.
• Add the assistant's response parts to the message history.
• Check stop_reason:
• If stop_reason is NOT equal to "tool use" → Break the loop (the conversation is complete).
• If stop_reason IS "tool use" → Proceed to tool execution.
• Execute the necessary tools using run_tools with the requested tool parts.
• Add the resulting tool output as a user message back into the history.
• Conversation Flow: This loop ensures that the conversation continues until the assistant provides a final response that does not request any further tool calls.

Takeaways

• Multi-turn conversations require a loop to manage the back-and-forth between the model and external tools.
• The stop_reason returned by the API is critical for determining if the conversation requires further tool execution or if it is complete.
• The chat function must be updated to return a structured dictionary containing parts, stop_reason, and a consolidated text.
• The conversation loop proceeds to tool execution only if the stop_reason is "tool use"; otherwise, the loop breaks.