[Bug]: Function calling in groupchat does not work

[LUK3ARK]

Describe the bug

I picked out the RAG groupchat example from the notebook and made several variations of it. I have been able to get it working if i chat only with the retrieve agent but when I initiate a conversation with the group manager, it suggests a function call and then fails.

When i run the agentchat_groupchat_RAG.ipynb example it suggests to call the retrieve function and then fails with this issue:

raise self._make_status_error_from_response(err.response) from None

openai.BadRequestError: Error code: 400 - {'error': {'message': "None is not of type 'array' - 'messages.2.tool_calls'", 'type': 'invalid_request_error', 'param': None, 'code': None}}

I get this error with MULTIPLE different functions.

I have ben debugging this for a while and my current conclusion that this bug comes from the groupchat manager itself. I tried hardcoding it so that an empty list is provided instead of None but they it says the message is still short.

I know that OpenAI is deprecating function calling and wanted to know if this is expected behaviour or if i am missing something vital?

Again - seems to only come when a manager is in the middle of the conversation, but have been able to see other agents call functions within a chat.

Steps to reproduce

Copy the agentchat_groupchat_RAG.ipynb and run it

Expected Behavior

When using the call rag function it should call the retrieve content function correctly

I expect to be able to reproduce the same results as the examples out of the box.

Screenshots and logs

lib/python3.11/site-packages/openai/_base_client.py", line 960, in _request raise self._make_status_error_from_response(err.response) from None openai.BadRequestError: Error code: 400 - {'error': {'message': "None is not of type 'array' - 'messages.2.tool_calls'", 'type': 'invalid_request_error', 'param': None, 'code': None}}

Additional Information

consistent in all versions of pyautogen down to 2.0.4, then it kind of works but other issues start tangling.

[LUK3ARK]

update: from testing, this seems to only occur if there is more than one user proxy in a group. When i have only one user proxy in a groupchat, function calling works correctly.

[thinkall]

@LUK3ARK I tried the sample notebook, it didn't raise any issue. https://github.com/microsoft/autogen/blob/fix_1440/notebook/agentchat_groupchat_RAG.ipynb

Could you share your notebook? Thanks.

When i run the agentchat_groupchat_RAG.ipynb example it suggests to call the retrieve function and then fails with this issue:

[LUK3ARK]

here is the script i run:

config_list = autogen.config_list_from_json(
    env_or_file="OAI_CONFIG_LIST",
    filter_dict={
        "model": ["gpt-3.5-turbo", "gpt-35-turbo", "gpt-3.5-turbo-0613", "gpt-4", "gpt4", "gpt-4-32k"],
    },
)

if __name__ == "__main__":

    llm_config = {
        "timeout": 60,
        "cache_seed": 42,
        "config_list": config_list,
        "temperature": 0,
    }

    # autogen.ChatCompletion.start_logging()

    def termination_msg(x):
        return isinstance(x, dict) and "TERMINATE" == str(x.get("content", ""))[-9:].upper()

    boss = autogen.UserProxyAgent(
        name="Boss",
        is_termination_msg=termination_msg,
        human_input_mode="NEVER",
        system_message="The boss who ask questions and give tasks.",
        code_execution_config=False,  # we don't want to execute code in this case.
        default_auto_reply="Reply `TERMINATE` if the task is done.",
    )

    boss_aid = RetrieveUserProxyAgent(
        name="Boss_Assistant",
        is_termination_msg=termination_msg,
        system_message="Assistant who has extra content retrieval power for solving difficult problems.",
        human_input_mode="NEVER",
        max_consecutive_auto_reply=3,
        retrieve_config={
            "task": "code",
            "docs_path": "https://raw.githubusercontent.com/microsoft/FLAML/main/website/docs/Examples/Integrate%20-%20Spark.md",
            "chunk_token_size": 1000,
            "model": config_list[0]["model"],
            "client": chromadb.PersistentClient(path="/tmp/chromadb"),
            "collection_name": "groupchat",
            "get_or_create": True,
        },
        code_execution_config=False,  # we don't want to execute code in this case.
    )

    coder = autogen.AssistantAgent(
        name="Senior_Python_Engineer",
        is_termination_msg=termination_msg,
        system_message="You are a senior python engineer. Reply `TERMINATE` in the end when everything is done.",
        llm_config=llm_config,
    )

    pm = autogen.AssistantAgent(
        name="Product_Manager",
        is_termination_msg=termination_msg,
        system_message="You are a product manager. Reply `TERMINATE` in the end when everything is done.",
        llm_config=llm_config,
    )

    reviewer = autogen.AssistantAgent(
        name="Code_Reviewer",
        is_termination_msg=termination_msg,
        system_message="You are a code reviewer. Reply `TERMINATE` in the end when everything is done.",
        llm_config=llm_config,
    )

    PROBLEM = "How to use spark for parallel training in FLAML? Give me sample code."

    def _reset_agents():
        boss.reset()
        boss_aid.reset()
        coder.reset()
        pm.reset()
        reviewer.reset()

    def rag_chat():
        _reset_agents()
        groupchat = autogen.GroupChat(
            agents=[boss_aid, coder, pm, reviewer], messages=[], max_round=12, speaker_selection_method="round_robin"
        )
        manager = autogen.GroupChatManager(groupchat=groupchat, llm_config=llm_config)

        # Start chatting with boss_aid as this is the user proxy agent.
        boss_aid.initiate_chat(
            manager,
            problem=PROBLEM,
            n_results=3,
        )

    def norag_chat():
        _reset_agents()
        groupchat = autogen.GroupChat(
            agents=[boss, coder, pm, reviewer],
            messages=[],
            max_round=12,
            speaker_selection_method="auto",
            allow_repeat_speaker=False,
        )
        manager = autogen.GroupChatManager(groupchat=groupchat, llm_config=llm_config)

        # Start chatting with the boss as this is the user proxy agent.
        boss.initiate_chat(
            manager,
            message=PROBLEM,
        )

    def call_rag_chat():
        _reset_agents()

        # In this case, we will have multiple user proxy agents and we don't initiate the chat
        # with RAG user proxy agent.
        # In order to use RAG user proxy agent, we need to wrap RAG agents in a function and call
        # it from other agents.
        def retrieve_content(message, n_results=3):
            boss_aid.n_results = n_results  # Set the number of results to be retrieved.
            # Check if we need to update the context.
            update_context_case1, update_context_case2 = boss_aid._check_update_context(message)
            if (update_context_case1 or update_context_case2) and boss_aid.update_context:
                boss_aid.problem = message if not hasattr(boss_aid, "problem") else boss_aid.problem
                _, ret_msg = boss_aid._generate_retrieve_user_reply(message)
            else:
                ret_msg = boss_aid.generate_init_message(message, n_results=n_results)
            return ret_msg if ret_msg else message

        boss_aid.human_input_mode = "NEVER"  # Disable human input for boss_aid since it only retrieves content.

        llm_config = {
            "functions": [
                {
                    "name": "retrieve_content",
                    "description": "retrieve content for code generation and question answering.",
                    "parameters": {
                        "type": "object",
                        "properties": {
                            "message": {
                                "type": "string",
                                "description": "Refined message which keeps the original meaning and can be used to retrieve content for code generation and question answering.",
                            }
                        },
                        "required": ["message"],
                    },
                },
            ],
            "config_list": config_list,
            "timeout": 60,
            "cache_seed": 42,
        }

        for agent in [coder, pm, reviewer]:
            # update llm_config for assistant agents.
            agent.llm_config.update(llm_config)

        for agent in [boss, coder, pm, reviewer]:
            # register functions for all agents.
            agent.register_function(
                function_map={
                    "retrieve_content": retrieve_content,
                }
            )

        groupchat = autogen.GroupChat(
            agents=[boss, coder, pm, reviewer],
            messages=[],
            max_round=12,
            speaker_selection_method="random",
            allow_repeat_speaker=False,
        )

        manager_llm_config = llm_config.copy()
        manager_llm_config.pop("functions")
        manager = autogen.GroupChatManager(groupchat=groupchat, llm_config=manager_llm_config)

        # Start chatting with the boss as this is the user proxy agent.
        boss.initiate_chat(
            manager,
            message=PROBLEM,
        )

    call_rag_chat()

Here is the execution result:

Details

Boss (to chat_manager): How to use spark for parallel training in FLAML? Give me sample code. -------------------------------------------------------------------------------- INFO:httpx:HTTP Request: POST https://api.openai.com/v1/chat/completions "HTTP/1.1 200 OK" Product_Manager (to chat_manager): To use Spark for parallel training in FLAML, you need to follow these steps: 1. Install FLAML and Apache Spark on your machine or cluster. 2. Import the necessary libraries: ```python from flaml import AutoML from pyspark.sql import SparkSession ``` 3. Create a SparkSession: ```python spark = SparkSession.builder \ .appName("FLAML with Spark") \ .getOrCreate() ``` 4. Load your data into a Spark DataFrame: ```python data = spark.read.format("csv").option("header", "true").load("path_to_your_data.csv") ``` 5. Initialize FLAML with Spark: ```python automl = AutoML() automl.initialize(spark=spark) ``` 6. Define the search space and other settings for FLAML: ```python settings = { "time_budget": 60, # total running time in seconds "metric": 'accuracy', # primary metric for optimization "task": 'classification', # task type "n_jobs": -1, # number of parallel jobs, -1 means using all available cores "log_file_name": 'flaml.log', # flaml log file } ``` 7. Start the FLAML training process: ```python automl.fit(data, **settings) ``` 8. Get the best model found by FLAML: ```python best_model = automl.best_model ``` 9. Make predictions using the best model: ```python predictions = best_model.predict(data) ``` 10. Terminate the SparkSession: ```python spark.stop() ``` That's it! You have successfully used Spark for parallel training in FLAML. Remember to adjust the settings and data loading code according to your specific use case. TERMINATE --------------------------------------------------------------------------------

-> i have tried with gpt-3.5-turbo-0613 and gpt-4 -> i have tried with export AUTOGEN_USE_DOCKER=0 and export AUTOGEN_USE_DOCKER=1 and both seem to not be able to use the function calling now. The issue before was that a function call was being suggested but then not able to be acted upon due to tool_calls being None.

Now however, it seems i just made the script again and i am not able to get it to reproduce even making the function call in the first place.

I have been able to get my usecase working so this is not a burning issue anymore but i had to go a different way around, I am still unable to just copy the setup and run it.

[xianzhisheng]

@LUK3ARK I had the same problem, how did you fix it?

[xianzhisheng]

I rolled back the version to 0.2.3 and the issue stopped happening

[ray30087]

How can I solve the problem with "file not found error". I'm using the notebook code as well. But I don't know how to deal with it.

[sonichi]

@thinkall could you take a note of this thread and inform the users in your RAG refactor roadmap?

[thinkall]

How can I solve the problem with "file not found error". I'm using the notebook code as well. But I don't know how to deal with it.

You need to either put your config file in the folder of the notebook or specifiy the file_location=<direcotry of the file OAI_CONFIG_LIST> to the function config_list_from_json.