You can use Honcho to manage user context around LLM frameworks like LangChain. First, import the appropriate packages:

from honcho import Honcho

from langchain_openai import ChatOpenAI
from langchain_core.prompts import ChatPromptTemplate, MessagesPlaceholder
from langchain_core.output_parsers import StrOutputParser
from langchain_core.messages import AIMessage, HumanMessage

from uuid import uuid4   # for generating random app and user names
from dotenv import load_dotenv   # for loading in LLM API keys

load_dotenv()   # assumes you have a .env file with OPENAI_API_KEY defined

Next let’s instantiate our Honcho client:

app_name = str(uuid4())

honcho = Honcho(app_name=app_name)
honcho.initialize()
user_name = str(uuid4())
user = honcho.create_or_get_user(user_name) # create or get user

Then we can define our chain using the LangChain Expression Language (LCEL):

prompt = ChatPromptTemplate.from_messages([
    ("system", "You are a helpful assistant."),
    MessagesPlaceholder(variable_name="chat_history"),
    ("user", "{input}")
])
model = ChatOpenAI(model="gpt-3.5-turbo")
output_parser = StrOutputParser()

chain = prompt | model | output_parser

Honcho returns lists of Message objects when queried using a built-in method like get_messages(), so a quick utility function is needed to change the list format to message objects LangChain expects:

def langchain_message_converter(messages: List):
    new_messages = []
    for message in messages:
        if message.is_user:
            new_messages.append(HumanMessage(content=message.content))
        else:
            new_messages.append(AIMessage(content=message.content))
    return new_messages

Now we can structure Honcho calls around our LLM inference:

sessions = list(user.get_sessions_generator(location_id))  # args come from application logic
session = sessions[0]   # most recent session for user
history = list(session.get_messages_generator())
chat_history = langchain_message_converter(history)   # convert messages for LangChain
inp = "Here's a user message!"
session.create_message(is_user=True, content=inp)

response = await chain.ainvoke({"chat_history": chat_history, "input": inp})

session.create_message(is_user=False, content=response)

Here we query messages from a user’s session using Honcho and construct a chat history object to send to the LLM alongside our immediate user input. Once the LLM has responded, we can add that to Honcho!

DiscordSimple Memory