GLM4 API server providing services, The parameter passed when LangChain calls the tool does not include the “query”.

[Jimmy-L99]

Checked other resources

• I added a very descriptive title to this question.
• I searched the LangChain documentation with the integrated search.
• I used the GitHub search to find a similar question and didn't find it.

Commit to Help

• I commit to help with one of those options 👆

Example Code

Part of the code is as follows:
'''python
llm_end_point_url = "my_end_point_url"
llm = ChatOpenAI(
    api_key=api_key, 
    temperature=0.5, 
    max_tokens=4096,
    model ="glm4v-9b", 
    base_url=llm_end_point_url
)
vector_store = Milvus(...)
retriever = vector_store.as_retriever(search_type="similarity", search_kwargs={"k": 3})

from langchain.tools.retriever import create_retriever_tool
tool = create_retriever_tool(
    retriever,
    "荔枝知识增强检索",
    "能够获取关于荔枝的专业知识，并对和荔枝相关的专业问题进行回答",
)
tools = [tool]

from langgraph.prebuilt import create_react_agent
from langchain_core.messages import HumanMessage
agent_executor = create_react_agent(llm, tools, debug=True)

query = "荔枝品种有哪些"
for s in agent_executor.invoke({"messages": [HumanMessage(content=query)]}):
    print(s)
    print("----") 
'''

> The following is the debug log:
> First llm:

'''
DEBUG    | __main__:create_chat_completion:417 - ==== request ====
{'messages': [ChatMessage(role='user', content='荔枝品种有哪些', image=None, function_call=None, tool_calls=None)], 'temperature': 0.5, 'top_p': 0.8, 'max_tokens': 4096, 'echo': False, 'stream': False, 'repetition_penalty': 1.1, 'tools': [{'type': 'function', 'function': {'name': '荔枝知识增强检索', 'description': '能够获取关于荔枝的专业知识，并对和荔枝相关的专业问题进行回答', 'parameters': {'type': 'object', 'properties': {'query': {'description': 'query to look up in retriever', 'type': 'string'}}, 'required': ['query']}}}], 'tool_choice': None}

==== message ====
role='assistant' content=None image=None function_call=None tool_calls=[ChatCompletionMessageToolCall(index=0, id='call_fkhWyvuFaxotK0owP8u4UGtm', function=FunctionCall(name='荔枝知识增强检索', arguments='{"function_name": "荔枝知识增强检索", "parameters": {"query": "请问有哪些荔枝品种？"}}'), type='function')]
'''

> Second llm:

'''
==== request ====
{'messages': [ChatMessage(role='user', content='荔枝品种有哪些', image=None, function_call=None, tool_calls=None), ChatMessage(role='assistant', content=None, image=None, function_call=None, tool_calls=[ChatCompletionMessageToolCall(index=0, id='call_fkhWyvuFaxotK0owP8u4UGtm', function=FunctionCall(name='荔枝知识增强检索', arguments='{"function_name": "\\u8354\\u679d\\u77e5\\u8bc6\\u589e\\u5f3a\\u68c0\\u7d22", "parameters": {"query": "\\u8bf7\\u95ee\\u6709\\u54ea\\u4e9b\\u8354\\u679d\\u54c1\\u79cd\\uff1f"}}'), type='function')]), ChatMessage(role='tool', content="Error: ValidationError(model='RetrieverInput', errors=[{'loc': ('query',), 'msg': 'field required', 'type': 'value_error.missing'}])\n Please fix your mistakes.", image=None, function_call=None, tool_calls=None)], 'temperature': 0.5, 'top_p': 0.8, 'max_tokens': 4096, 'echo': False, 'stream': False, 'repetition_penalty': 1.1, 'tools': [{'type': 'function', 'function': {'name': '荔枝知识增强检索', 'description': '能够获取关于荔枝的专业知识，并对和荔枝相关的专业问题进行回答', 'parameters': {'type': 'object', 'properties': {'query': {'description': 'query to look up in retriever', 'type': 'string'}}, 'required': ['query']}}}], 'tool_choice': None}
'''

Description

I use milvus database as vector store to build RAG, follow the langchain tutorials to create tool, but when I use tool and it return some error.
I notice that in the first llm the glm4 does give the argument:

function=FunctionCall(name='荔枝知识增强检索', arguments='{"function_name": "荔枝知识增强检索", "parameters": {"query": "请问有哪些荔枝品种？"}}'), type='function')

I think it seems right, this is the right argument format? After the agent, langchain gave some wrong input in the second llm, that cause the (Error: ValidationError)
Why the second call request have Unicode character：

arguments='{"function_name": "\\u8354\\u679d\\u77e5\\u8bc6\\u589e\\u5f3a\\u68c0\\u7d22", "parameters": {"query": "\\u8bf7\\u95ee\\u6709\\u54ea\\u4e9b\\u8354\\u679d\\u54c1\\u79cd\\uff1f"}}'), type='function')])

and tool message with error.

ChatMessage(role='tool', content="Error: ValidationError(model='RetrieverInput', errors=[{'loc': ('query',), 'msg': 'field required', 'type': 'value_error.missing'}])\n Please fix your mistakes."

What am I supposed to do to fix this problem and use tool successfully?

Thank you very much for your suggestions and solutions

System Info

python--3.11.9
langchain--0.2.11
langgraph--0.2.14
langchain-openai--0.1.22

[dosubot[bot]]

@2500035435 Hello! I'm here to help you with your GitHub issues. I can help you solve bugs, answer questions, and become a contributor.

The issue you're encountering with the GLM4 API server and the ValidationError for the 'query' parameter is due to the Unicode characters in the second call request. This happens because the api_request_chain and api_answer_chain in LangChain expect specific input variables, and if these variables are not correctly formatted or contain unexpected characters, it can lead to validation issues.

In your case, the 'query' parameter contains Unicode characters that are not properly handled, causing the ValidationError. The RetrieverInput model requires the query field to be a string and properly formatted.

To resolve this, ensure that the 'query' parameter is properly sanitized and encoded to handle Unicode characters. Here is a relevant part of the code that validates the input variables:

class RetrieverInput(BaseModel):
    """Input to the retriever."""
    query: str = Field(description="query to look up in retriever")

If the 'query' parameter contains Unicode characters, it must be properly handled to prevent validation errors. You can sanitize and encode the 'query' parameter before passing it to the tool.

Additionally, ensure that the api_request_chain and api_answer_chain are correctly set up to handle the expected input variables:

@root_validator(pre=False, skip_on_failure=True)
def validate_api_request_prompt(cls, values: Dict) -> Dict:
    input_vars = values["api_request_chain"].prompt.input_variables
    expected_vars = {"question", "api_docs"}
    if set(input_vars) != expected_vars:
        raise ValueError(f"Input variables should be {expected_vars}, got {input_vars}")
    return values

@root_validator(pre=False, skip_on_failure=True)
def validate_api_answer_prompt(cls, values: Dict) -> Dict:
    input_vars = values["api_answer_chain"].prompt.input_variables
    expected_vars = {"question", "api_docs", "api_url", "api_response"}
    if set(input_vars) != expected_vars:
        raise ValueError(f"Input variables should be {expected_vars}, got {input_vars}")
    return values

Ensure that the 'query' parameter is properly formatted and encoded to handle Unicode characters to prevent these validation errors ^[1][2].

To continue talking to Dosu, mention @dosu.

[Jimmy-L99]

As the code see，

agent_executor = create_react_agent(llm, tools, debug=True)
query = "荔枝品种有哪些"
agent_executor.invoke({"messages": [HumanMessage(content=query)]})

The query parameter just Chinese character, and I just invoke the query to agent_executor, the second call request sender is langchain agent, what cause that, is the first call output argument format wrong?

[dosubot[bot]]

The issue you're encountering with the ValidationError for the 'query' parameter is likely due to how the create_react_agent function and the agent_executor.invoke method handle the input messages and the transformation of the query parameter.

The create_react_agent function processes a list of messages, including the query, and does not directly transform the query parameter. However, it works with the list of messages that include the query ^[1].

The ValidationError you're seeing indicates that the 'query' parameter is missing or not correctly formatted in the second call. This could be due to how the function_call arguments are handled and parsed, especially with non-ASCII characters like Chinese.

The JsonOutputFunctionsParser class in LangChain handles the parsing of function_call arguments and has a strict attribute that, when set to False, allows for non-JSON-compliant strings, including handling Unicode characters ^[2].

To resolve this issue, ensure that the query parameter is correctly formatted and encoded to handle Unicode characters. You can set the strict attribute to False in the JsonOutputFunctionsParser to allow for non-JSON-compliant strings:

class JsonOutputFunctionsParser(BaseCumulativeTransformOutputParser[Any]):
    strict: bool = False  # Allow non-JSON-compliant strings
    args_only: bool = True

    def parse_result(self, result: List[Generation], *, partial: bool = False) -> Any:
        # Parsing logic here

Additionally, ensure that the query parameter is properly validated as a string when creating the retriever tool:

class RetrieverInput(BaseModel):
    query: str = Field(description="query to look up in retriever")

By ensuring proper handling and encoding of the query parameter, you can prevent the ValidationError and successfully use the tool.

To continue talking to Dosu, mention @dosu.

[Jimmy-L99]

This time I use english as the tool name, tool description and query.
`tool = create_retriever_tool(
retriever,
"Litchi_knowledge_base",
"Be able to acquire professional knowledge about litchi and answer professional questions related to litchi",
)

tools = [tool]
agent_executor = create_react_agent(llm, tools, debug=True)

query = "What are the litchi varieties"
for s in agent_executor.stream({"messages": [HumanMessage(content=query)]}):
print(s)
print("----")'

The second request as follow:

==== request ==== {'messages': [ChatMessage(role='user', content='What are the litchi varieties', image=None, function_call=None, tool_calls=None), ChatMessage(role='assistant', content=None, image=None, function_call=None, tool_calls=[ChatCompletionMessageToolCall(index=0, id='call_SUkIhw8QCOp9oclx3vUHh4iP', function=FunctionCall(name='Litchi_knowledge_base', arguments='{"function_name": "Litchi_knowledge_base", "parameters": {"query": "What are the different varieties of litchis?"}}'), type='function')]), ChatMessage(role='tool', content="Error: ValidationError(model='RetrieverInput', errors=[{'loc': ('query',), 'msg': 'field required', 'type': 'value_error.missing'}])\n Please fix your mistakes.", image=None, function_call=None, tool_calls=None)], 'temperature': 0.5, 'top_p': 0.8, 'max_tokens': 4096, 'echo': False, 'stream': False, 'repetition_penalty': 1.1, 'tools': [{'type': 'function', 'function': {'name': 'Litchi_knowledge_base', 'description': 'Be able to acquire professional knowledge about litchi and answer professional questions related to litchi', 'parameters': {'type': 'object', 'properties': {'query': {'description': 'query to look up in retriever', 'type': 'string'}}, 'required': ['query']}}}], 'tool_choice': None}

The ChatMessage of tool still have error:
ChatMessage(role='tool', content="Error: ValidationError(model='RetrieverInput', errors=[{'loc': ('query',), 'msg': 'field required', 'type': 'value_error.missing'}])\n Please fix your mistakes."