Why does chaining LangChain components with the | operator cause a ValueError with ChatOpenAI?

I'm developing a chatbot using LangChain with OpenAI's gpt-3.5-turbo model. When I use the LLMChain abstraction, everything works as expected, but when I chain components with the | operator, I encounter the following error:

PLAINTEXTCopied
ValueError: Invalid input type <class 'dict'>. Must be a PromptValue, str, or list of BaseMessages.

This is my code snippet:

PYTHONCopied
def chatbot(input_user_message):
    # creating a prompt template
    chat_prompt = ChatPromptTemplate.from_messages(
        [
            ("system", "You are a helpful AI assistant."),
            MessagesPlaceholder(variable_name="history_messages"),
            ("human", "{input_user_message}"),
        ]
    )

    # initializing OpenAI Chat model
    llm = ChatOpenAI(model="gpt-3.5-turbo", temperature=0.7)

    trimmer = trim_messages(
        max_tokens=100,
        strategy="last",
        token_counter=llm,
        # Usually, we want to keep the SystemMessage
        include_system=True,
        # start_on="human" makes sure we produce a valid chat history
        start_on="human",
    )

    def get_session_history(session_id):
        if session_id not in st.session_state.store:
            st.session_state.store[session_id] = ChatMessageHistory()
        else:
            st.session_state.store[session_id].messages = trimmer.invoke(st.session_state.store[session_id].messages)
        return st.session_state.store[session_id]

    # Initializing the output parser
    output_parser = StrOutputParser()

    # Creating an LLMChain with the prompt and memory
    #conversation_chain = LLMChain(
     #   llm=llm,
     #   prompt=chat_prompt,
     #   output_parser=output_parser,  # To parse the LLM's response into a string for display
        # verbose=True,  # Displays detailed logs for debugging
    #)
    conversation_chain = llm | chat_prompt | output_parser

    model_with_memory = RunnableWithMessageHistory(
        conversation_chain,
        get_session_history,
        input_messages_key="input_user_message",
        history_messages_key="history_messages",
    )

    session_id = "1234"
    # config = {"configurable": {"session_id": session_id}}

    response = model_with_memory.invoke(
        {"input_user_message": input_user_message},
        {"configurable": {"session_id": session_id}},
    )
    print(response)
    return response["text"]

The error suggests that one of the chain components is receiving a dictionary instead of the expected prompt type (a PromptValue, string, or list of BaseMessages).

I understand that LLMChain automatically maps dictionary inputs to the right format for each placeholder, while using the | operator requires sequentially passing a valid type from one component to the next.

My questions are:

1. Why does the chaining order with the | operator require precise input types for each component?
2. How can I adjust my chain or inputs so that the pipe operator works correctly with ChatOpenAI and ChatPromptTemplate?

I have tried ensuring proper extraction of placeholders by using building blocks like MessagesPlaceholder, but the issue persists. Any guidance or best practices when chaining Runnable objects in LangChain would be greatly appreciated.

I am using python==3.12.7, langchain==0.3.9, openai==1.55.0, and langchain-openai==0.2.10.