Issue: How can we combine ConversationalRetrievalChain with openai ChatCompletion

[apurv101]

Issue you'd like to raise.

I'm trying to chat with a PDF using ConversationalRetrievalChain.

embeddings = OpenAIEmbeddings()
vectordb = Chroma(embedding_function=embeddings, persist_directory=directory)
qa_chain = ConversationalRetrievalChain.from_llm(ChatOpenAI(model_name="gpt-3.5-turbo", temperature=0.3), vectordb.as_retriever(), memory=memory)
answer = (qa_chain({"question": query}))

It works perfectly as it gives the answer from the documents. But it can't alter the tone or even convert the answer into another language when prompted.
How can we change the tone like we do in openai ChatCompletion:

    response = openai.ChatCompletion.create(
    model="gpt-3.5-turbo",
    messages=[
            {"role": "system", "content": "You are a helpful and friendly chatbot who converts text to a very friendly tone."},
            {"role": "user", "content": f"{final_answer}"}
        ]
    )

such that it answers from the doc but also converts it according to some given prompt. Right now I have to pass the received output from ConversationalRetrievalChain in above code in order to modify the tone. Is this kind of functionality missing in ConversationalRetrievalChain?

Suggestion:

No response

[dosubot]

Answer generated by a 🤖

Answer

Hello,

Thank you for your question. Based on the information provided and the current state of the LangChain codebase, the ConversationalRetrievalChain class does not support tone modification or language translation. Its primary function is to retrieve relevant documents based on a given question and chat history, and it does not have any functionality for modifying the tone of the conversation.

However, you can achieve the desired functionality by combining the ConversationalRetrievalChain with OpenAI's ChatCompletion. Here's a simple way to do it:

# Get the answer from the documents
embeddings = OpenAIEmbeddings()
vectordb = Chroma(embedding_function=embeddings, persist_directory=directory)
qa_chain = ConversationalRetrievalChain.from_llm(ChatOpenAI(model_name="gpt-3.5-turbo", temperature=0.3), vectordb.as_retriever(), memory=memory)
answer = (qa_chain({"question": query}))

# Modify the tone using OpenAI's ChatCompletion
response = openai.ChatCompletion.create(
    model="gpt-3.5-turbo",
    messages=[
        {"role": "system", "content": "You are a helpful and friendly chatbot who converts text to a very friendly tone."},
        {"role": "user", "content": f"{answer}"}
    ]
)

In this way, you first retrieve the answer from the documents using ConversationalRetrievalChain, and then pass the answer to OpenAI's ChatCompletion to modify the tone.

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

Sources

• libs/langchain/langchain/chains/conversational_retrieval/base.py

---

This response is meant to be useful, save you time, and share context. 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.

[apurv101]

Added a section in documentation answering this

https://python.langchain.com/docs/use_cases/question_answering/how_to/chat_vector_db#using-a-custom-prompt-for-condensing-the-question