Ollama tool calls not working via openai proxy only when using langgraph

[StreetLamb]

Checked other resources

• I added a very descriptive title to this issue.
• I searched the LangGraph/LangChain documentation with the integrated search.
• I used the GitHub search to find a similar question and didn't find it.
• I am sure that this is a bug in LangGraph/LangChain rather than my code.
• I am sure this is better as an issue rather than a GitHub discussion, since this is a LangGraph bug and not a design question.

Example Code

from typing import Literal

from langchain_core.messages import HumanMessage
from langchain_core.tools import tool
from langchain_openai import ChatOpenAI
from langgraph.checkpoint import MemorySaver
from langgraph.graph import END, MessagesState, StateGraph
from langgraph.prebuilt import ToolNode
import asyncio


@tool
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."


tools = [search]

tool_node = ToolNode(tools)

model = ChatOpenAI(
    model="llama3.1", base_url="http://localhost:11434/v1", temperature=0
).bind_tools(tools)


def should_continue(state: MessagesState) -> Literal["tools", END]:
    messages = state["messages"]
    last_message = messages[-1]
    if last_message.tool_calls:
        return "tools"
    return END


async def call_model(state: MessagesState, config):
    messages = state["messages"]
    response = await model.ainvoke(messages, config)
    return {"messages": [response]}


workflow = StateGraph(MessagesState)

workflow.add_node("agent", call_model)
workflow.add_node("tools", tool_node)

workflow.set_entry_point("agent")

workflow.add_conditional_edges(
    "agent",
    should_continue,
)

workflow.add_edge("tools", "agent")

checkpointer = MemorySaver()

app = workflow.compile(checkpointer=checkpointer)

async def test():
    async for event in app.astream_events(
        {"messages": [HumanMessage(content="what is the weather in sf")]},
        version="v1",
        config={"configurable": {"thread_id": 42}},
    ):
        print(event)


asyncio.run(test())

Error Message and Stack Trace (if applicable)

No response

Description

I want to invoke a tool-calling compatible Ollama model through ChatOpenAI proxy. However, using the code above, the model does not properly tool call:

{'event': 'on_chain_end', 'data': {'output': {'messages': [HumanMessage(content='what is the weather in sf', id='f7017ae4-b2d0-49e3-b939-69738686368b'), AIMessage(content='{"name": "search", "parameters": {"query": "sf weather"}}', response_metadata={'finish_reason': 'stop', 'model_name': 'llama3.1', 'system_fingerprint': 'fp_ollama'}, id='run-6a214185-27ba-4505-9cc2-574f20d04909')]}}, 'run_id': 'b2aeba64-38b0-447a-b7de-eefff49e3555', 'name': 'LangGraph', 'tags': [], 'metadata': {'thread_id': 43}, 'parent_ids': []}

However, the behaviour is different when using just langchain:

import asyncio

import openai
from langchain_core.messages import HumanMessage
from langchain_core.prompts import ChatPromptTemplate, MessagesPlaceholder
from langchain_core.tools import tool
from langchain_openai import ChatOpenAI

@tool
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."


model = ChatOpenAI(model="llama3.1", base_url="http://localhost:11434/v1", temperature=0)

model_with_tools = model.bind_tools([search])

async def test():
    prompt = ChatPromptTemplate.from_messages(
        [
            MessagesPlaceholder(variable_name="messages"),
        ]
    )
    chain = prompt | model_with_tools
    response = await chain.ainvoke(
        {"messages": [HumanMessage(content="What is the weather like in sf")]}
    )
    return response


response = asyncio.run(test())
print(response)

This way the model correctly utilise a tool call:

content='' additional_kwargs={'tool_calls': [{'id': 'call_xncx3ycn', 'function': {'arguments': '{"query":"sf weather"}', 'name': 'search'}, 'type': 'function'}]} response_metadata={'token_usage': {'completion_tokens': 17, 'prompt_tokens': 148, 'total_tokens': 165}, 'model_name': 'llama3.1', 'system_fingerprint': 'fp_ollama', 'finish_reason': 'stop', 'logprobs': None} id='run-63a9efd2-6619-448b-9a89-476f45cfb5c8-0' tool_calls=[{'name': 'search', 'args': {'query': 'sf weather'}, 'id': 'call_xncx3ycn', 'type': 'tool_call'}] usage_metadata={'input_tokens': 148, 'output_tokens': 17, 'total_tokens': 165}

System Info

langchain==0.2.7
langchain-anthropic==0.1.20
langchain-cohere==0.1.5
langchain-community==0.2.7
langchain-core==0.2.21
langchain-google-genai==1.0.5
langchain-ollama==0.1.0
langchain-openai==0.1.17
langchain-qdrant==0.1.1
langchain-text-splitters==0.2.0
langchain-weaviate==0.0.1.post1

platform: mac silicon
python version: Python 3.12.2

[hinthornw]

Oh interesting. How many times did you run both versions? Is it reliably different in both contexts? And then you've confirmed the ollama versions are the same in both scenarios?

[StreetLamb]

Hi @hinthornw , I ran them multiple times (and once more just now just to be sure) and the behaviour is consistent. Tested both on ollama v0.3.0.

[StreetLamb]

It seems to be cause by astream_events method. Using langgraph with astream works:

...
async def test():
    async for event in app.astream(
        {"messages": [HumanMessage(content="what is the weather in sf")]},
        config={"configurable": {"thread_id": 42}},
    ):
        print(event)


asyncio.run(test())

{'agent': {'messages': [AIMessage(content='', additional_kwargs={'tool_calls': [{'id': 'call_rs3ykbgl', 'function': {'arguments': '{"query":"sf weather"}', 'name': 'search'}, 'type': 'function'}]}, response_metadata={'token_usage': {'completion_tokens': 17, 'prompt_tokens': 147, 'total_tokens': 164}, 'model_name': 'llama3.1', 'system_fingerprint': 'fp_ollama', 'finish_reason': 'stop', 'logprobs': None}, id='run-3175464b-50ce-4fa7-afc7-73fe14cf92ee-0', tool_calls=[{'name': 'search', 'args': {'query': 'sf weather'}, 'id': 'call_rs3ykbgl', 'type': 'tool_call'}], usage_metadata={'input_tokens': 147, 'output_tokens': 17, 'total_tokens': 164})]}}
{'tools': {'messages': [ToolMessage(content="It's 60 degrees and foggy.", name='search', tool_call_id='call_rs3ykbgl')]}}
{'agent': {'messages': [AIMessage(content='Based on the tool call response, I can format an answer to your original question:\n\nThe current weather in San Francisco (SF) is 60 degrees with fog.', response_metadata={'token_usage': {'completion_tokens': 34, 'prompt_tokens': 86, 'total_tokens': 120}, 'model_name': 'llama3.1', 'system_fingerprint': 'fp_ollama', 'finish_reason': 'stop', 'logprobs': None}, id='run-0c048f94-63ae-45d1-9808-4083dd65ec0d-0', usage_metadata={'input_tokens': 86, 'output_tokens': 34, 'total_tokens': 120})]}}