How can I make a chain that includes a System Message, a Retriever and a Memory all at once?

[ZaidHani]

Checked other resources

• I added a very descriptive title to this question.
• I searched the LangChain documentation with the integrated search.
• I used the GitHub search to find a similar question and didn't find it.

Commit to Help

• I commit to help with one of those options 👆

Example Code

executor = initialize_agent(
    agent=AgentType.OPENAI_FUNCTIONS,
    # the tools parameter includes the function RetrievalQA.from_chain_type, it is how I query the data from the vector database
    tools=tools,
    llm=llm,
    memory=memory,
    system_message=system_message,
    agent_kwargs={
        "system_message": system_message,
        "extra_prompt_messages": [MessagesPlaceholder(variable_name="memory")]
    },
    verbose=False,
    handle_parsing_errors=True
)

Description

I'm trying to build an equivalent to this agent code using chains, I want to use chains because they are much faster than agents, but so far I'm facing a problem. I can't seem to find a way to insert a system message, a retriever and a memory all at once into a chain, each solution I've found solve this problem by providing 2 of these previous features and excluding the last one, I've found someone who built a chain and used a system message and a retriever but no memory, and I've found someone who used a retriever and a memory but with no system message... so far I've opted to use an agent because it has all of these features combined, is there a way to build the same code using only chains. thank you.

System Info

langchain==0.1.4
langchain-community==0.0.16
langchain-core==0.1.17
langchain-openai==0.0.5

platform widows

python version 3.12.1

[dosubot[bot]]

🤖

Hey @ZaidHani! Great to see you back with another interesting question. Let's dive into it.

Based on the context provided, it seems you want to create a chain that includes a System Message, a Retriever, and a Memory all at once. While there isn't a built-in way to do this in LangChain, you can extend the BaseConversationalRetrievalChain class to incorporate a system message into the retrieval process. Here's an example:

from langchain_core.language_models import BaseLanguageModel
from langchain_core.retrievers import BaseRetriever
from langchain_core.messages import SystemMessage
from langchain.chains.conversational_retrieval import BaseConversationalRetrievalChain

class MyChain(BaseConversationalRetrievalChain):
    def __init__(self, system_message: SystemMessage, *args, **kwargs):
        super().__init__(*args, **kwargs)
        self.system_message = system_message

    def _get_docs(
        self,
        question: str,
        inputs: Dict[str, Any],
        *,
        run_manager: CallbackManagerForChainRun,
    ) -> List[Document]:
        # Incorporate the system message into the retrieval process
        question = f"{self.system_message.content} {question}"
        return super()._get_docs(question, inputs, run_manager=run_manager)

# Usage
llm = BaseLanguageModel()  # Replace with your actual language model
retriever = BaseRetriever()  # Replace with your actual retriever
system_message = SystemMessage(content="System message content")

chain = MyChain(system_message, llm=llm, retriever=retriever)

In this example, a new class MyChain is created that extends BaseConversationalRetrievalChain. The system_message is passed to the constructor and stored as an instance variable. Then, in the _get_docs method, the content of the system_message is prepended to the question before calling the parent class's _get_docs method. This way, the system message is incorporated into the retrieval process.

Please replace BaseLanguageModel and BaseRetriever with your actual language model and retriever.

You can find more details in the source code of the BaseConversationalRetrievalChain class in the LangChain codebase: libs/langchain/langchain/chains/conversational_retrieval/base.py

As for the memory, you can use the ConversationBufferWindowMemory class to store the chat history. This class is designed to store a window of the most recent messages in a conversation. You can pass an instance of this class to the memory parameter of the initialize_agent function.

Here's an example:

from langchain.chains.conversation.memory import ConversationBufferWindowMemory

# Create the memory
memory = ConversationBufferWindowMemory(
    memory_key="chat_history", return_messages=True
)

# Pass the memory to the initialize_agent function
executor = initialize_agent(
    agent=AgentType.OPENAI_FUNCTIONS,
    tools=tools,
    llm=llm,
    memory=memory,
    system_message=system_message,
    agent_kwargs={
        "system_message": system_message,
        "extra_prompt_messages": [MessagesPlaceholder(variable_name="memory")]
    },
    verbose=False,
    handle_parsing_errors=True
)

In this example, ConversationBufferWindowMemory is used to create a memory that stores the chat history. The memory_key parameter is set to "chat_history", and return_messages is set to True to return the messages as instances of BaseMessage. This memory is then passed to the initialize_agent function.

You can find more details in the source code of the ConversationBufferWindowMemory class in the LangChain codebase: libs/langchain/langchain/chains/conversation/memory.py

I hope this helps! If you have any other questions, feel free to ask.

Sources

• No support for ConversationalRetrievalChain with passing in memory and system message
• Question: How can I include SystemMessage with RetrievalQA
• Declaring and Open AI Functions Agent with memory and system message
• libs/langchain/langchain/chains/conversational_retrieval/base.py
• libs/langchain/langchain/agents/initialize.py

About Dosu

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.

[arthurdeka]

I'm also interested on the answer to this. Did you managed to make it work @ZaidHani ?