AIMessage enforces ToolCall model incompatible with openai

[yildirimgoks]

When you provide a value for the 'tool_calls' argument of the AIMessage, root_validator tries to parse the values in it to a ToolCall object using the langchain_core.messages.tool.tool_call function. Although instancing ToolCall by calling it directly with OpenAI compatible tool_calls model is possible, the tool_call() function expects a specific set of variables to build the ToolCall, and does not accept extra keyword variables. This enforces a very specific data model when tool_calls is provided with AIMessage and makes it impossible to create an AIMessage object with the desired tool_calls formatting.

Is this intended or could this be a bug? IIs there a different way of creating an AI message with tool_calls that I'm missing?

Example code:

from langchain.messages import AIMessage, ToolCall

# OpenAI compatible tool call format
tool_call = {
   "id":"tc1",
   "type":"function",
   "function":{
      "name":"SomeTool",
      "arguments":"{\"some_argument\": \"some_value\"}"
   }
}

# line below works fine
ToolCall(**tool_call)

# line below returns an error
AIMessage(content="", tool_calls=[ToolCall(**tool_call)])

# error message:
# pydantic.v1.error_wrappers.ValidationError: 1 validation error for AIMessage
# tool_call() got an unexpected keyword argument 'function' (type=type_error)

[dosubot[bot]]

Hey @yildirimgoks! 👋 I'm here to help you with bugs, questions, and becoming a contributor. I'm Dosu, a bot designed to assist you while you wait for a human maintainer.

There are similar discussions and an issue related to the problem you're encountering:

1. AIMessage enforces ToolCall model incompatible with openai - This discussion is ongoing and was last updated on November 5, 2024 ^[1].

2. 'error': {'message': "Missing parameter 'tool_call_id': messages with role 'tool' must have a 'tool_call_id' - This discussion is also unsolved and was last updated on June 6, 2024 ^[2].

3. OllamaFunctions returning ValueError: tool_calls missing from AIMessage: {message} - This issue is closed but did not have a solution or comments provided. It was last updated on October 15, 2024 ^[3].

The problem seems to stem from the differences in the expected format for tool calls between OpenAI and LangChain. LangChain's ToolCall expects a specific format that does not align with the OpenAI format. In LangChain, the ToolCall should be structured with a name, args as a dictionary, and optionally an id, rather than using a function object with JSON strings for arguments ^[4].

To continue talking to Dosu, mention @dosu.

[yildirimgoks]

Yes but ToolCall(**tool_call) works with OpenAI compatible parameters, so it doesn't seem intentional to enforce the name-args-id strucutre for a ToolCall. Is there a way to create an AIMessage instance by bypassing the extra validation layer AIMessage does on top of ToolCall? @dosu

[dosubot[bot]]

The issue arises because the AIMessage class in LangChain enforces a specific structure for ToolCall objects, which may not align with the OpenAI format. While ToolCall(**tool_call) works because it directly instantiates a ToolCall object, the AIMessage class performs additional validation that expects a different format.

To bypass this validation, you could consider creating a custom wrapper or subclass that adjusts the AIMessage validation logic to accept the OpenAI format. Alternatively, you might need to manually convert the OpenAI tool call format to the expected LangChain format before passing it to AIMessage. This involves restructuring the tool call to use a name, args as a dictionary, and optionally an id, instead of a function object with JSON strings for arguments ^[1][2].

To continue talking to Dosu, mention @dosu.