Using HugginFaceChat instead of ChatOpenAI in load_chat_planner

[RamishSiddiqui]

Hi community,

I am trying to Plan_and_Execute agent to use multiple agents in a single query. is this possible ? Here is what i have come up with.

`
memory = ConversationBufferMemory(
memory_key="chat_history",
return_messages=True,
output_key='answer',
chat_memory=chat_history
)

tools = [
    Tool(
        name="Conversation_LLM",
        func=LLMChain(
            llm=llm,
            prompt=prompt_chat,
            verbose=True,
            memory=memory
        ).run,
        description="Useful for having a simple chat."
    ),
    Tool(
        name="Calculator_LLM",
        func=LLMMathChain.from_llm(
            llm=llm,
            prompt=prompt_math,
            verbose=True,
            memory=memory
        ).run,
        description="Useful for solving math problems."
    )
]

model = llm #Using huggingface here instead of ChatOpenAI from langchains.chat_models
planner = load_chat_planner(model)
executor = load_agent_executor(model, tools, verbose=True)
agent = PlanAndExecute(
    planner=planner,
    executor=executor,
    verbose=True,
    memory=memory,
)

result = agent(query["text"])

`

I am getting this error.

`

Entering new PlanAndExecute chain...
Traceback (most recent call last):
File "D:\Seer App\seer_main\seer.py", line 44, in
result = chain(query["text"])
File "C:\Users\Ramish\anaconda3\envs\Seer App\lib\site-packages\langchain\chains\base.py", line 258, in call
raise e
File "C:\Users\Ramish\anaconda3\envs\Seer App\lib\site-packages\langchain\chains\base.py", line 252, in call
self._call(inputs, run_manager=run_manager)
File "C:\Users\Ramish\anaconda3\envs\Seer App\lib\site-packages\langchain\experimental\plan_and_execute\agent_executor.py", line 68, in _call
return {self.output_key: self.step_container.get_final_response()}
File "C:\Users\Ramish\anaconda3\envs\Seer App\lib\site-packages\langchain\experimental\plan_and_execute\schema.py", line 55, in get_final_response
return self.steps[-1][1].response
IndexError: list index out of range
steps=[]
Process finished with exit code 1
`

[dosubot[bot]]

🤖

Hello,

From the error message you provided, it seems like the PlanAndExecute agent is not able to find any steps to execute, hence the IndexError: list index out of range error. This could be due to the planner not being able to generate any plans for the given query.

The PlanAndExecute agent in LangChain is designed to handle multiple agents by iterating over the steps planned by the planner. For each step, it prepares the inputs, which include the previous steps, the current step, and the objective. It then calls the executor's step method to execute the current step. The response from the executor is then added to the step container. This process is repeated for each step in the plan.

Here is the relevant code that shows how it handles multiple agents:

for step in plan.steps:
    _new_inputs = {
        "previous_steps": self.step_container,
        "current_step": step,
        "objective": inputs[self.input_key],
    }
    new_inputs = {**_new_inputs, **inputs}
    response = self.executor.step(
        new_inputs,
        callbacks=run_manager.get_child() if run_manager else None,
    )
    self.step_container.add_step(step, response)

In your case, it seems like the planner is not able to generate any steps for the given query. This could be due to the model not being able to generate any plans, or the tools not being able to handle the given query. You might want to check the implementation of your load_chat_planner function and the tools you're using.

As for using HuggingFaceChat instead of ChatOpenAI, you should be able to replace the model used in the load_chat_planner and load_agent_executor functions with any model that is compatible with the LangChain framework. However, you need to ensure that the model is able to generate plans and execute them correctly.

I hope this helps! Let me know if you have any other questions.

Sources

• libs/langchain/tests/unit_tests/agents/test_agent.py
• libs/experimental/langchain_experimental/plan_and_execute/agent_executor.py

---

This response is meant to be useful and save you time. It is not meant to be a precise solution, but rather a starting point for your own research.

Help me be more useful! Please leave a 👍 if this is helpful and 👎 if it is irrelevant.

If you want to continue the conversation, start your reply with @dosu-bot.

[RamishSiddiqui]

Can you provide an example where plan_and_execute uses huggingface chat model instead of an OpenAI one.

[dosubot[bot]]

🤖

Hello,

From the error message you provided, it seems like the PlanAndExecute agent is not generating any steps for the given query, which is causing the IndexError: list index out of range error. This could be due to the planner not being compatible with the HuggingFace chat model.

The PlanAndExecute agent in LangChain uses a planner and an executor to plan and execute a chain of steps. The planner is responsible for planning the steps, and the executor is responsible for executing each step. If the planner does not generate any steps, the executor will not have any steps to execute, which will cause an IndexError.

To use the HuggingFace chat model with the PlanAndExecute agent, you need to ensure that the planner and executor are compatible with the HuggingFace chat model. Here's an example of how you can do this:

from langchain.chat_models import ChatHuggingFace  # Import HuggingFace chat model

# Replace ChatOpenAI with ChatHuggingFace
model = ChatHuggingFace(temperature=0)

planner = load_chat_planner(model)
executor = load_agent_executor(model, tools, verbose=True)
agent = PlanAndExecute(planner=planner, executor=executor, verbose=True)

Then you can run the agent as before:

agent.run("Who is Leo DiCaprio's girlfriend? What is her current age raised to the 0.43 power?")

Please note that the ChatHuggingFace class is a placeholder and you need to replace it with the actual class name of the HuggingFace chat model in the LangChain framework. The ChatHuggingFace class should have similar methods and properties as the ChatOpenAI class for this code to work.

If you are still encountering the same error, it could be due to a bug in the LangChain framework. In that case, please provide more details about the query you are trying to run and the version of LangChain you are using, so we can investigate further.

I hope this helps! Let me know if you have any other questions.

Sources

• docs/snippets/modules/agents/agent_types/plan_and_execute.mdx
• libs/experimental/langchain_experimental/plan_and_execute/agent_executor.py

---

This response is meant to be useful and save you time. It is not meant to be a precise solution, but rather a starting point for your own research.

Help me be more useful! Please leave a 👍 if this is helpful and 👎 if it is irrelevant.

If you want to continue the conversation, start your reply with @dosu-bot.

[RamishSiddiqui]

There is no class in
from langchain.chat_models import ChatHuggingFace

Are you hallucinating ?