tools require a filled description

[efriis]

Generally langchain tools default to having an empty string as a description. Some (maybe all) bedrock models require a filled description, which affects the ease of use of with_structured_output(method="function_calling") and bind_tools with ChatBedrock

In particular, this impacts model interoperability, where people are trying their tools with different models, and the bedrock versions simply fail.

Some ideas:

• populate the description with the same as the tool name by default in bedrock to avoid the error
• raise a more descriptive error to help people populate docstrings on their tool function/basemodel/etc

I'm probably more in favor of auto populating for models this impacts

cc @hinthornw

[3coins]

cc @patilvishal0597

[patilvishal0597]

Hello, I tried to recreate this problem with a tool below. I noticed that langchain auto-populates docstrings in tool description field if description is not provided. I have commented the docstrings in the below example:

class WikiInputs(BaseModel):
    # """Inputs to the wikipedia tool."""

    query: str = Field(
        description="query to look up in Wikipedia, should be 3 or less words"
    )

@tool("search-tool", args_schema=WikiInputs, return_direct=True)
def search(query: str) -> str:
    # """Look up things online."""
    return "LangChain"

tools = [search]

I have the following code that is executed next:

def main():
    """Initializing Bedrock Chat for testing ChatBedrock calls"""
    llm = getChatBedrock().bind_tools(tools=tools)

    
    """llm Invoke/converse call"""
    response = llm.invoke(getInputMessage())
    print(response)

I get the following stack trace:

Traceback (most recent call last):
  File "/Users/vishal/Desktop/langchain-aws/langchain-aws/libs/aws/tool_use_example.py", line 18, in <module>
    @tool("search-tool", args_schema=WikiInputs, return_direct=True)
     ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/Users/vishal/Library/Caches/pypoetry/virtualenvs/langchain-aws-uKC4YU62-py3.12/lib/python3.12/site-packages/langchain_core/tools/convert.py", line 249, in _tool_factory
    return StructuredTool.from_function(
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/Users/vishal/Library/Caches/pypoetry/virtualenvs/langchain-aws-uKC4YU62-py3.12/lib/python3.12/site-packages/langchain_core/tools/structured.py", line 190, in from_function
    raise ValueError(msg)
ValueError: Function must have a docstring if description not provided.

Can I request you to provide an example of tool use where the description is an empty string? This can be helpful to debug further. Thanks!

[3coins]

@patilvishal0597
@tool decorator does require a description as required in the function used as a tool, this is defined in the docs, but you can also bind simple functions as tools (without the @tool decorator). I think the issue raised here by @efriis is related to that use case. Here is a sample code that replicates the issue.

from pydantic import BaseModel, Field
from langchain_aws import ChatBedrockConverse


def search(query: str) -> str:
    return "LangChain"


def main():
    llm = ChatBedrockConverse(
        model="anthropic.claude-3-5-haiku-20241022-v1:0"
    )
    llm_with_tools = llm.bind_tools([search])
    llm_with_tools.invoke("Hello!")


if __name__ == "__main__":
    main()

Here is the exception raised by the Bedrock API.

    raise ParamValidationError(report=report.generate_report())
botocore.exceptions.ParamValidationError: Parameter validation failed:
Invalid length for parameter toolConfig.tools[0].toolSpec.description, value: 0, valid min length: 1