Help with the tool arguments being sent

[avishwa23]

Hi, I was running this crewai example with the latest version of crewai (crewai==0.76.2, crewai-tools==0.13.2, langchain==0.3.4) and I have internet searching function as a tool taking in a query argument (seen in the image below).

The issue is that my LLM (OpenAI model (gpt-4o-2024-08-06)) is always returning argument as name instead of query in its response (e.g., {name: 'ACTUAL_SEARCH_QUERY_FROM_LLM'}). I am not sure if the LLM is misinterpreting the name argument in the tool arguments section (highlighted in the picture) as the function argument.

Can anyone help with (1) understanding where this tool definition in the prompt is coming from and (2) how to resolve this error (to have llm output QUERY instead of NAME)?
Thanks.

[zholti]

Experiencing the same issue with crewai==0.76.9, pydantic==2.10.0.b1.

[zholti]

I believe that this behavior is a bug in the Agent class in agent.py, specifically in the _render_text_description_and_args function.

In this function, the input args for the tools are built from all fields on the Pydantic type, as opposed to the args_schema fields only:

args_schema = str(tool.model_fields)

Instead, I’d gather the args like this:

args_schema = {key: val['type'] for key, val in tool.args.items()}

Am I missing something or can we fix this as described?

[tonykipkemboi]

Hi @avishwa23,
Are you running this example code?

[avishwa23]

Hi @tonykipkemboi , almost similar, the only difference being the way I am setting the model. I am setting the model name as an environment variable using OPENAI_MODEL_NAME which I guess CrewAI looks for. I am using the latest gpt-4o.

And the crewAI library version (0.76.2) is the latest (as of writing the original thread) as well.

[avishwa23]

@zholti

Instead, I’d gather the args like this:

I think you are right…At least the _render_text_description_and_args method is expected to do so (based on its docstrings, seen below). Do you think its a minor bug in the code @tonykipkemboi ?
Thanks.