app/rag.py

import os
import streamlit as st
from langsmith import traceable
from langchain_community.llms import Replicate
from langchain.memory import ConversationBufferWindowMemory
from langchain.chains import ConversationalRetrievalChain
from langchain_community.document_transformers import LongContextReorder
from app.prompts import sahabat_prompt


def reorder_embedding(docs):
    """Reorder documents for long context retrieval."""
    reordering = LongContextReorder()
    return reordering.transform_documents(docs)


@traceable(name="Create RAG Conversational Chain")
def create_conversational_chain(vector_store):
    """Create a Conversational Retrieval Chain for RAG with tuned parameters for lower latency."""
    llm = Replicate(
        model="fauzi3007/sahabat-ai-replicate:c3fc398f441379bd3fb6a4498950f9302aa75b7a95e76978a689ceb5c4b4bf09",
        model_kwargs={"temperature": 0.1, "top_p": 0.9, "max_new_tokens": 3000}
    )
    memory = ConversationBufferWindowMemory(
        k=6,
        memory_key="chat_history",
        return_messages=True,
        output_key="answer"
    )
    chain = ConversationalRetrievalChain.from_llm(
        llm,
        retriever=vector_store.as_retriever(search_type="mmr", search_kwargs={"k": 6, "fetch_k": 20}),
        combine_docs_chain_kwargs={"prompt": sahabat_prompt},
        return_source_documents=True,
        memory=memory,
    )
    return chain


def get_rag_chain(vector_store):
    """Return a Conversational Retrieval Chain for external use."""
    return create_conversational_chain(vector_store)