Transition

SSethThunder

Hello everyone. Hope everyones doing well.

I'm currently doing my internship and I've been tasked to do a chatbot which I've been trying to do for the past 2-3 months using LangChain but I struggling a lot understanding and implementing several things from their docs so decided to switch to LlamaIndex and try it out. Could someone help me out transition from what I had in LangChain to LlamaIndex if someone understood both?

I'm currently doing RAG that has custom prompt, memory, and retrieves from a DirectoryLoader. I've also implemented Constitutional AI to prevent Prompt Injection/Hacking.

If someone could help me out, I'd highly appreciate it as I'm currently left with 2 weeks to finish it. I can show my LangChain code as well if it helps.

18 comments

WWhiteFang_Jr

Sure, I'll try my best to help you in the transition.

SSethThunder

Hello Fang

SSethThunder

Thanks for replying

SSethThunder

Do you want me to share my code?

WWhiteFang_Jr

Yes

WWhiteFang_Jr

That would help me to understand your code and give better suggestions

SSethThunder

Sure 1sec

WWhiteFang_Jr

hey

WWhiteFang_Jr

created something from your code

WWhiteFang_Jr

Had one doubt

WWhiteFang_Jr

you have two bots one to give response and other to verify if the first response is correct or not?

SSethThunder

Thats 1 issue I'm just trying to fix but the 2nd response has constitutional AI that prevents prompt hacking

SSethThunder

Is there a way to implement it using Llama index?

WWhiteFang_Jr

I think for this
response_2 = qa_chain({"retrieval_answer": llm_response["result"], "question": question})

you can tackle this with a OpenAI agent as well.

SSethThunder

But arent agents known to have bad accuracy?

SSethThunder

Thats why I've been avoiding them

WWhiteFang_Jr

You can use the LLM directly then

Plain Text

llm_response = chat_engine.chat(question)
response_2 = llm.complete("Define instructions for validating the answer:\n ADD_RETRIEVED_RESPONSE_HERE\n\nADD_QUESTION")

WWhiteFang_Jr

OpenAI agent have been pretty good. specially if you are using GPT-4-Preview model

Add a reply

Find answers from the community

Transition