Exercise#3 LangChain memory

Objective

Learn to use the LangChain ConversationSummaryMemory class.

  • Decide on LLM used for Chatbot responses
  • Decide on LLM used for summarizing the history

Checkout final bot on HuggingFace space : acloudfan/chatbot-langchain-memory

Hints

1. Make a copy of the Streamlit chatbot

  • Create a folder (langchain-chatbot-interface)
  • Copy Streamlt/streamlit-chatbot-interface/app.py to new folder
  • Make changes to the app.py to accopalish the task

exercise-chatbot-interface

2. Memory class instance

  • Checkout the API documentation for ConversationSummaryMemory
  • Instead of using Streamlit session_state, you will use the ConversationSummaryClass
  • Create a function to get the summarization LLM
def  get_summarization_llm():
     model = 'gpt-3.5-turbo-0125'
     return ChatOpenAI(model=model, openai_api_key=openai_api_key)
  • Create instance of the ConversationSummaryMemory class and add it to StreamLit session state.
if 'MEMORY' not in st.session_state:
    memory = ConversationSummaryMemory(
        llm = get_summarization_llm(),
        human_prefix='user',
        ai_prefix = 'assistant',
        return_messages=True
    )
    # add to the session
    st.session_state['MEMORY'] = memory

3. Create the ConversationChain

def get_llm_chain():
    memory = st.session_state['MEMORY']
    conversation = ConversationChain(
        llm=get_llm(),
        memory=memory
    )
    return conversation

4. Dump the summary conversation history

  • Show the conversation history on the interface
st.divider()
st.subheader('Context/Summary:')

# Print the state of the buffer
st.session_state['MEMORY'].buffer

5. To stop Streamlit from processing

st.stop()

Solution

Solution uses the OpenAI gpt3.5-turbo model for both the chatbot responses and for context summarization. You may replace them with any model of your choice. The solution script is available in the project repository.

exercise-chatbot-interface