[LangGraph + Ollama] Agent using local model (qwen2.5) returns AIMessage(content='') even when tool responds correctly

[matiasdev30]

[LangGraph + Ollama] Agent using local model (qwen2.5) returns AIMessage(content='') even when tool responds correctly

I’m using create_react_agent from langgraph.prebuilt with a local model served via Ollama (qwen2.5), and the agent consistently returns an AIMessage with an empty content field — even though the tool returns a valid string.

Code

from langgraph.prebuilt import create_react_agent
from langchain_ollama import ChatOllama

model = ChatOllama(model="qwen2.5")

def search(query: str):
    """Call to surf the web."""
    if "sf" in query.lower() or "san francisco" in query.lower():
        return "It's 60 degrees and foggy."
    return "It's 90 degrees and sunny."

agent = create_react_agent(model=model, tools=[search])

response = agent.invoke(
    {},
    {"messages": [{"role": "user", "content": "what is the weather in sf"}]}
)
print(response)

Output

{
  'messages': [
    AIMessage(
      content='', 
      additional_kwargs={}, 
      response_metadata={
        'model': 'qwen2.5',
        'created_at': '2025-04-24T09:13:29.983043Z',
        'done': True,
        'done_reason': 'load',
        'total_duration': None,
        'load_duration': None,
        'prompt_eval_count': None,
        'prompt_eval_duration': None,
        'eval_count': None,
        'eval_duration': None,
        'model_name': 'qwen2.5'
      },
      id='run-6a897b3a-1971-437b-8a98-95f06bef3f56-0'
    )
  ]
}

As shown above, the agent responds with an empty string, even though the search() tool clearly returns "It's 60 degrees and foggy.".

Has anyone seen this behavior? Could it be an issue with qwen2.5, langgraph.prebuilt, the Ollama config, or maybe a mismatch somewhere between them?

Any insight appreciated.
``

[andreygithub123]

The problem is in how you pass the messages dict to the .invoke method. The .invoke() expects two arguments - an input dict and a config dict. Here is the method signature: def invoke(self, input: dict, config: dict = None) -> dict .When you pass messages as you did, you actually pass an empty dict for input and a config dict with messages :

response = agent.invoke(
    {}, # input dict
    {"messages": [{"role": "user", "content": "what is the weather in sf"}]} # config dict
)

The key here is to simply pass the messages as the input dict and it will probably work.

response = agent.invoke(
    {"messages": [{"role": "user", "content": "what is the weather in sf"}]}, # input dict
    # empty or no config dict
)
print(response)

or specify the keyword arguments :

response = agent.invoke(
    config={},
    input={"messages": [{"role": "user", "content": "what is the weather in sf"}]}
    
)

Also make sure that the chosen LLM from ollama supports tool calling. As far as I know qwen2.5 supports it but it might be a good idea to switch to a more powerful LLM.