I’m creating a chatbot that is also context-aware using a vector retriever. This is accomplished using create_stuff_documents_chain, create_retrieval_chain, and RunnableWithMessageHistory.

I can get the chat history working with just the RedisChatMessageHistory function, but after a few interactions, the chat history becomes huge because it returns every single AI/Human interaction within the current conversation.

I had an idea that to pipe the RedisChatMessageHistory into the ConversationSummaryBufferMemory’s own chat_memory argument, and use the resultant output as the session history to keep it relatively short. I tried that with my code here, but it does not work (see error below):

question_answer_chain = create_stuff_documents_chain(llm, qa_prompt)

rag_chain = create_retrieval_chain(
    history_aware_retriever, question_answer_chain)

# Statefully manage chat history
def get_session_history(session_id: str) -> BaseChatMessageHistory:
    history = RedisChatMessageHistory(
        session_id,
        url="redis://:" + settings.passwd + "@redis",
        key_prefix=settings.prefix
    )

    return ConversationSummaryBufferMemory(
        llm=llm,
        memory_key="chat_history",
        input_key="input",
        max_token_limit=10,
        return_messages=False,
        chat_memory=history
    )

conversational_rag_chain = RunnableWithMessageHistory(
    rag_chain,
    get_session_history,
    input_messages_key="input",
    history_messages_key="chat_history",
    output_messages_key="answer"
)

# Stream the result
for s in conversational_rag_chain.stream(
    {"input": the_question},
    config={
        "configurable": {"session_id": conversation_session_id}
    }
):
    # print(s, flush=True)
    if ("answer" in s):
        yield s["answer"]

The error message:

AttributeError: ‘ConversationSummaryBufferMemory’ object has no attribute ‘messages’

If I remove the ConversationSummaryBufferMemory call and return the Redis message history byitself as below, then it all works great:

# Statefully manage chat history
def get_session_history(session_id: str) -> BaseChatMessageHistory:
    history = RedisChatMessageHistory(
        session_id,
        url="redis://:" + settings.passwd+ "@redis",
        key_prefix=settings.prefix
    )

    return history

Am I approaching this the wrong way? Is there a way to summarise the Redis chat history in a similar way to ConversationSummaryBufferMemory?

Thank you!