Why do workflow tools need to return JSON and TEXT two same results?

[swingchen01]

Self Checks

• I have searched for existing issues search for existing issues, including closed ones.
• I confirm that I am using English to submit this report (我已阅读并同意 Language Policy).
• [FOR CHINESE USERS] 请务必使用英文提交 Issue，否则会被关闭。谢谢！:）
• Please do not modify this template :) and fill in all the required fields.

1. Is this request related to a challenge you're experiencing? Tell me about your story.

Workflow tools return JSON and TEXT two same results, are there any special scenarios to consider?
This will cause the FunctionCallAgentRunner to provide the duplicate tool results to the LLM"s API, which will take up twice the number of tokens.

2. Additional context or comments

If two same result is no special purpose, this logic will bring more performance consumption, more token consumption and fees, and more LLM context occupation

3. Can you help us with this feature?

• I am interested in contributing to this feature.

[dosubot[bot]]

I found two similar discussions regarding the issue of workflow tools returning duplicate results in JSON and TEXT formats:

1. Why do workflow tools need to return two identical results? One is JSON, and the other is TEXT, which is currently unsolved and was last updated on November 9, 2024 ^[1].

2. How to resolve the issue where the agent tool invocation returns duplicate response content, which is also unsolved and was last updated on September 9, 2024 ^[2].

These discussions might provide additional insights or context regarding the rationale behind this design choice and any potential solutions or workarounds.

To continue talking to Dosu, mention @dosu.

[crazywoola]

If you want to use it in the agent, the agent will take all things it produces, such as json message and text message, however you can choose the answer you want in workflow. There is a workaround, you can publish the workflow as a tool, and return text or json message only, then you won't get the duplicated message I think.

[swingchen01]

If you want to use it in the agent, the agent will take all things it produces, such as json message and text message, however you can choose the answer you want in workflow. There is a workaround, you can publish the workflow as a tool, and return text or json message only, then you won't get the duplicated message I think.

The two results are generated in WorkflowTool._invoke()' when the tool is called, and it is already called after being published as a tool, so the solution you mentioned cannot solve the problem. Refer to WorkflowTool._invoke()` method 76/77 lines of code:

result.append(self.create_text_message(json.dumps(outputs, ensure_ascii=False)))
result.append(self.create_json_message(outputs))

Then refer to ToolEngine._convert_tool_response_to_str() method:

def _convert_tool_response_to_str(tool_response: list[ToolInvokeMessage]) -> str:
    """
    Handle tool response
    """
    result = ""
    for response in tool_response:
        if response.type == ToolInvokeMessage.MessageType.TEXT:
            result += response.message
        elif response.type == ToolInvokeMessage.MessageType.LINK:
            result += f"result link: {response.message}. please tell user to check it."
        elif (
            response.type == ToolInvokeMessage.MessageType.IMAGE_LINK
            or response.type == ToolInvokeMessage.MessageType.IMAGE
        ):
            result += (
                "image has been created and sent to user already, you do not need to create it,"
                " just tell the user to check it now."
            )
        elif response.type == ToolInvokeMessage.MessageType.JSON:
            result += f"tool response: {json.dumps(response.message, ensure_ascii=False)}."
        else:
            result += f"tool response: {response.message}."

    return result

This method generates twice the text content result. Finally, refer to line 259 of the FunctionalCallAgentRunner. run() method to generate a ToolPromptMessage object with twice the result:

if tool_response["tool_response"] is not None:
    self._current_thoughts.append(
        ToolPromptMessage(
            content=tool_response["tool_response"],
            tool_call_id=tool_call_id,
            name=tool_call_name,
        )
    )