search

langchain-ai / langchain

🦜🔗 Build context-aware reasoning applications
https://python.langchain.com
MIT License
94.92k stars 15.37k forks source link

Issue with Ollama Function – Only Agent Response, No Tool Calls #28110

Open Billian-D opened 2 days ago

Billian-D commented 2 days ago

Checked other resources

Example Code

from langchain_core.messages import AIMessage
from langgraph.graph import END, StateGraph
from typing import TypedDict
from langchain_core.tools import tool
from langchain_experimental.llms.ollama_functions import OllamaFunctions
from langchain_core.pydantic_v1 import BaseModel, Field
from langchain.prompts import PromptTemplate
from termcolor import colored
import json
# Import tools from tools.py
from tools import get_current_weather, get_system_time

# using OllamaFunctions from experimental because it supports function binding with llms
model = OllamaFunctions(
    base_url="http://localhost:11436",
    model="llama3.1",
    format="json"
    )

model_with_tools = model.bind_tools(
    tools=[get_current_weather, get_system_time],
)

tool_mapping = {
    'get_current_weather': get_current_weather,
    'get_system_time': get_system_time,
}

# Define Agent Prompt template for llama3
agent_request_generator_prompt = PromptTemplate(
    template=
    """
    <|begin_of_text|>
    <|start_header_id|>system<|end_header_id|>
        You are a Smart Agent. 
        You are a master at understanding what a customer wants.
        You evaluate every request and utilize available tools if you have to.
    <|eot_id|>
    <|start_header_id|>user<|end_header_id|>
    Conduct a comprehensive analysis of the request provided\

    USER REQUEST:\n\n {initial_request} \n\n

    <|eot_id|>
    <|start_header_id|>assistant<|end_header_id|>
    """,
    input_variables=["initial_request"],
)

agent_request_generator = agent_request_generator_prompt | model_with_tools
# result = agent_request_generator.invoke({"initial_request": "What is the weather in woodbury in MN?"})
# print(result)
# input("...")

# Pydantic Schema for structured response
class Evaluation(BaseModel):
    result: bool = Field(description="True or False", required=True)

# Prompt template llama3
category_generator_prompt = PromptTemplate(
    template=
    """
    <|begin_of_text|>
    <|start_header_id|>system<|end_header_id|>
        You are a Smart Router Agent. You are a master at reviewing whether the original question that customer asked was answered in the tool response.
        You understand the context and question below and return your answer in JSON.
    <|eot_id|>
    <|start_header_id|>user<|end_header_id|>
    CONTEXT: Conduct a comprehensive analysis of the Initial Request from user and Tool Response and route the request into boolean true or false:
        True - used when INITIAL REQUEST appears to be answered by TOOL RESPONSE. \
        False - used when INITIAL REQUEST is not answered by TOOL RESPONSE or when TOOL RESPONSE is empty \

            Output either True or False \
            eg:
            'True' \n\n
    INITIAL REQUEST:\n\n {research_question} \n\n
    TOOL RESPONSE:\n\n {tool_response} \n\n

    JSON:
    <|eot_id|>
    <|start_header_id|>assistant<|end_header_id|>
    """,
 input_variables=["research_question", "tool_response"],
)

structured_llm = model.with_structured_output(Evaluation)
category_generator = category_generator_prompt | structured_llm

# result = category_generator.invoke({"research_question": "What is the weather in woodbury in MN?", "tool_response":"65C, Sunny"})
# print(result)
# input("...")

class AgentState(TypedDict):
    research_question: str
    tool_response: str
    agent_response: AIMessage
    agent_call_count: int = 0
    tool_call_count: int = 0

def agent(state: AgentState):
    print(colored("STATE at agent start:", "magenta"), colored(state, "cyan"))
    input("Paused ... Hit Enter to Execute Agent Logic...")
    last_ai_message = agent_request_generator.invoke({"initial_request": state["research_question"]})
    state["agent_call_count"] += 1
    #append the response to the agent_response list in the state
    if last_ai_message is not None:
        state["agent_response"] = last_ai_message 
        if last_ai_message.content is not None and last_ai_message.content != "" :
            state["tool_response"]=last_ai_message.content
    print(colored("STATE at agent end:", "magenta"), colored(state, "cyan"))
    input("Paused Hit Enter to go to Should Continue Logic...")    
    return state

def should_continue(state: AgentState):
    print(colored("STATE at should_continue start:", "magenta"), colored(state, "cyan"))
    input("Paused at Should Continue Start")
    print(colored("Evaluating whether the Question is Answered by the tool response or not... Please wait...", "red"))
    result = category_generator.invoke({"research_question": state["research_question"],
                                         "tool_response":state["tool_response"]
                                        })
    if isinstance(result, Evaluation):
        # Access the 'result' attribute from Evaluation
        print(colored("Is tool response good and should the flow go to END node? ", "cyan"), colored(result.result, "yellow"))
        input("Paused at Should Continue Mid")

        if result.result:  # If result is True
            print(colored("Return end", "red"))
            return "end"
        else:  # If result is False
            print(colored("Return continue", "green"))
            return "continue"
    else:
        print("Result is not an Evaluation instance, returning 'end' as default.")
        return "end"

def call_tool(state: AgentState):
    print(colored("STATE at call_tool start:", "magenta"), colored(state, "cyan"))
    input("Paused at call_tool Start")
    agent_response = state["agent_response"]

    if hasattr(agent_response, 'tool_calls') and len(agent_response.tool_calls) > 0:
        tool_call = agent_response.tool_calls[0]
        tool = tool_mapping[tool_call["name"].lower()]
        try:
            tool_output = tool.invoke(tool_call["args"])
            state["tool_call_count"] += 1
            print(colored("Tool output:", "magenta"), colored(tool_output, "green"))
            if tool_output is not None:
                state["tool_response"] = tool_output
        except Exception as e:
            print(f"Error invoking tool: {e}")
            # Handle the error or log it as needed
    else:
        print("No tool calls found in agent response.")
    print(colored("STATE at call_tool end:", "magenta"), colored(state, "cyan"))
    input("Paused at call_tool End")
    return state

workflow = StateGraph(AgentState)

workflow.add_node("agent", agent)
workflow.add_node("action", call_tool)

workflow.set_entry_point("agent")

workflow.add_conditional_edges(
    "agent",
    should_continue,
    {
        "continue": "action",
        "end": END,
    },
)
workflow.add_edge("action", "agent")

app = workflow.compile()

#helper method to visualize graph
def save_graph_to_file(runnable_graph, output_file_path):
    png_bytes = runnable_graph.get_graph().draw_mermaid_png()
    with open(output_file_path, 'wb') as file:
        file.write(png_bytes)

save_graph_to_file(app, "output-05.png")

# research_question = "What's the current system time?"
research_question = "Tell me a joke?"
# research_question = "How is the weather in Woodbury MN today?"
# research_question = "What is the cause of earthquakes?"

state : AgentState = {"research_question": research_question,
                      "tool_response": [] ,
                      "agent_response": [],
                      "agent_call_count": 0,
                      "tool_call_count": 0
                      }
result = app.invoke(state)
print("\n")
print(colored("FINAL STATE at end:", "magenta"), colored(result, "cyan"))

print(colored("FINAL RESPONSE at end:", "magenta"), colored(result["tool_response"], "cyan"))

Error Message and Stack Trace (if applicable)

STATE at agent start: {'research_question': 'Tell me a joke?', 'tool_response': [], 'agent_response': [], 'agent_call_count': 0, 'tool_call_count': 0}
Paused ... Hit Enter to Execute Agent Logic...
STATE at agent end: {'research_question': 'Tell me a joke?', 'tool_response': "Here's one: Why couldn't the bicycle stand up by itself? Because it was two-tired!", 'agent_response': AIMessage(content="Here's one: Why couldn't the bicycle stand up by itself? Because it was two-tired!", additional_kwargs={}, response_metadata={}, id='run-409504c1-afb2-49f7-8ace-bc08c89fdd33-0'), 'agent_call_count': 1, 'tool_call_count': 0}
Paused Hit Enter to go to Should Continue Logic...
STATE at should_continue start: {'research_question': 'Tell me a joke?', 'tool_response': "Here's one: Why couldn't the bicycle stand up by itself? Because it was two-tired!", 'agent_response': AIMessage(content="Here's one: Why couldn't the bicycle stand up by itself? Because it was two-tired!", additional_kwargs={}, response_metadata={}, id='run-409504c1-afb2-49f7-8ace-bc08c89fdd33-0'), 'agent_call_count': 1, 'tool_call_count': 0}
Paused at Should Continue Start
Evaluating whether the Question is Answered by the tool response or not... Please wait...
Traceback (most recent call last):
  File "/home/poc/agentic-workflow/langgraph-learning/tutorials/example_1.py", line 203, in <module>
    result = app.invoke(state)
             ^^^^^^^^^^^^^^^^^
  File "/home/miniconda3/lib/python3.11/site-packages/langgraph/pregel/__init__.py", line 1749, in invoke
    for chunk in self.stream(
  File "/home/miniconda3/lib/python3.11/site-packages/langgraph/pregel/__init__.py", line 1477, in stream
    for _ in runner.tick(
  File "/home/miniconda3/lib/python3.11/site-packages/langgraph/pregel/runner.py", line 58, in tick
    run_with_retry(t, retry_policy)
  File "/home/miniconda3/lib/python3.11/site-packages/langgraph/pregel/retry.py", line 29, in run_with_retry
    task.proc.invoke(task.input, config)
  File "/home/miniconda3/lib/python3.11/site-packages/langgraph/utils/runnable.py", line 412, in invoke
    input = context.run(step.invoke, input, config)
            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/miniconda3/lib/python3.11/site-packages/langgraph/utils/runnable.py", line 184, in invoke
    ret = context.run(self.func, input, **kwargs)
          ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/miniconda3/lib/python3.11/site-packages/langgraph/graph/graph.py", line 95, in _route
    result = self.path.invoke(value, config)
             ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/miniconda3/lib/python3.11/site-packages/langgraph/utils/runnable.py", line 176, in invoke
    ret = context.run(self.func, input, **kwargs)
          ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/poc/agentic-workflow/langgraph-learning/tutorials/example_1.py", line 121, in should_continue
    result = category_generator.invoke({"research_question": state["research_question"],
             ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/miniconda3/lib/python3.11/site-packages/langchain_core/runnables/base.py", line 3024, in invoke
    input = context.run(step.invoke, input, config)
            ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/miniconda3/lib/python3.11/site-packages/langchain_core/runnables/base.py", line 5354, in invoke
    return self.bound.invoke(
           ^^^^^^^^^^^^^^^^^^
  File "/home/miniconda3/lib/python3.11/site-packages/langchain_core/language_models/chat_models.py", line 286, in invoke
    self.generate_prompt(
  File "/home/miniconda3/lib/python3.11/site-packages/langchain_core/language_models/chat_models.py", line 786, in generate_prompt
    return self.generate(prompt_messages, stop=stop, callbacks=callbacks, **kwargs)
           ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/miniconda3/lib/python3.11/site-packages/langchain_core/language_models/chat_models.py", line 643, in generate
    raise e
  File "/home/miniconda3/lib/python3.11/site-packages/langchain_core/language_models/chat_models.py", line 633, in generate
    self._generate_with_cache(
  File "/home/miniconda3/lib/python3.11/site-packages/langchain_core/language_models/chat_models.py", line 851, in _generate_with_cache
    result = self._generate(
             ^^^^^^^^^^^^^^^
  File "/home/miniconda3/lib/python3.11/site-packages/langchain_experimental/llms/ollama_functions.py", line 305, in _generate
    functions = [convert_to_ollama_tool(fn) for fn in functions]
                ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/miniconda3/lib/python3.11/site-packages/langchain_experimental/llms/ollama_functions.py", line 305, in <listcomp>
    functions = [convert_to_ollama_tool(fn) for fn in functions]
                 ^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/home/miniconda3/lib/python3.11/site-packages/langchain_experimental/llms/ollama_functions.py", line 86, in convert_to_ollama_tool
    schema = tool.model_construct().model_json_schema()
             ^^^^^^^^^^^^^^^^^^^^
AttributeError: type object 'Evaluation' has no attribute 'model_construct'

Description

I have encountered an issue where the ollama function returns only the agent's response but does not invoke any associated tools, as expected. I have ensured that all required libraries and dependencies are properly installed, but this behavior persists.

System Info

"pip install langgraph httpx ipython graphviz langchain langchain_openai langchain_experimental langchain-ollama termcolor rich ollama openai docker"

platform (linux) python version 3.11

langgraph==0.2.46 langchain==0.2.17 langchain-experimental==0.0.65 langchain-core==0.2.43

BennisonDevadoss commented 2 days ago

I am also facing the same issue, Any update on it?