app.py

import gradio as gr
import os
from langchain_community.vectorstores import FAISS
from langchain_community.document_loaders import PyPDFLoader
from langchain.text_splitter import RecursiveCharacterTextSplitter
from langchain_community.vectorstores import Chroma
from langchain.chains import ConversationalRetrievalChain
from langchain_community.embeddings import HuggingFaceEmbeddings 
from langchain_community.llms import HuggingFacePipeline
from langchain.chains import ConversationChain
from langchain.memory import ConversationBufferMemory
from langchain_community.llms import HuggingFaceEndpoint
import torch

api_token = os.getenv("HF_TOKEN")

list_llm = ["microsoft/Phi-3-mini-4k-instruct", "mistralai/Mistral-7B-Instruct-v0.3"]  
list_llm_simple = [os.path.basename(llm) for llm in list_llm]

# Load and split PDF document
def load_doc(list_file_path, chunk_size, chunk_overlap):
    loaders = [PyPDFLoader(x) for x in list_file_path]
    pages = []
    for loader in loaders:
        pages.extend(loader.load())
    text_splitter = RecursiveCharacterTextSplitter(
        chunk_size=chunk_size, 
        chunk_overlap=chunk_overlap 
    )  
    doc_splits = text_splitter.split_documents(pages)
    return doc_splits

# Create vector database
def create_db(splits):
    embeddings = HuggingFaceEmbeddings()
    vectordb = FAISS.from_documents(splits, embeddings)
    return vectordb

# Initialize langchain LLM chain
def initialize_llmchain(llm_model, vector_db, progress=gr.Progress()):
    llm = HuggingFaceEndpoint(
        huggingfacehub_api_token=api_token,
        repo_id=llm_model, 
        temperature=0.1,
        max_new_tokens=2000,
        top_k=3,
    )
    
    memory = ConversationBufferMemory(
        memory_key="chat_history",
        output_key='answer',
        return_messages=True
    )

    retriever = vector_db.as_retriever()
    qa_chain = ConversationalRetrievalChain.from_llm(
        llm,
        retriever=retriever,
        chain_type="stuff", 
        memory=memory,
        return_source_documents=True,
        verbose=False,
    )
    return qa_chain

# Initialize database
def initialize_database(list_file_obj, chunk_size, chunk_overlap, progress=gr.Progress()):
    list_file_path = [x.name for x in list_file_obj if x is not None]
    doc_splits = load_doc(list_file_path, chunk_size, chunk_overlap)
    vector_db = create_db(doc_splits)
    if vector_db is None:
        print("Vector database creation failed")
    else:
        print("Embedding database created successfully")
    return vector_db, "Embedding database created!"

# Initialize LLM
def initialize_LLM(llm_option, vector_db, progress=gr.Progress()):
    if vector_db is None:
        print("Vector database is None")
        return None, "Failed to initialize RAG System: Vector database is None"
    
    llm_name = list_llm[llm_option]
    qa_chain = initialize_llmchain(llm_name, vector_db, progress)
    return qa_chain, "RAG System initialized!"

def format_chat_history(message, chat_history):
    formatted_chat_history = []
    for user_message, bot_message in chat_history:
        formatted_chat_history.append(f"User: {user_message}")
        formatted_chat_history.append(f"Assistant: {bot_message}")
    return formatted_chat_history

def conversation(qa_chain, message, history):
    formatted_chat_history = format_chat_history(message, history)
    response = qa_chain.invoke({"question": message, "chat_history": formatted_chat_history})
    response_answer = response["answer"]
    if response_answer.find("Helpful Answer:") != -1:
        response_answer = response_answer.split("Helpful Answer:")[-1]
    response_sources = response["source_documents"]
    response_source1 = response_sources[0].page_content.strip()
    response_source2 = response_sources[1].page_content.strip()
    response_source3 = response_sources[2].page_content.strip()
    response_source1_page = response_sources[0].metadata["page"] + 1
    response_source2_page = response_sources[1].metadata["page"] + 1
    response_source3_page = response_sources[2].metadata["page"] + 1
    new_history = history + [(message, response_answer)]
    return qa_chain, gr.update(value=""), new_history, response_source1, response_source1_page, response_source2, response_source2_page, response_source3, response_source3_page

def upload_file(file_obj):
    list_file_path = []
    for idx, file in enumerate(file_obj):
        file_path = file_obj.name
        list_file_path.append(file_path)
    return list_file_path

def demo():
    with gr.Blocks(theme=gr.themes.Default(primary_hue="green")) as demo:
        vector_db = gr.State()
        qa_chain = gr.State()
        gr.HTML("<center><h1>RAG System</h1><center>")
        gr.Markdown("""This App is designed to perform retrieval augmented generation (RAG) on PDF documents. \
        <b>Please do not upload confidential documents.</b>
        """)
        with gr.Row():
            with gr.Column(scale=86):
                gr.Markdown("<b>Step 1 - Upload PDF documents and Initialize the RAG system</b>")
                with gr.Row():
                    document = gr.Files(height=300, file_count="multiple", file_types=["pdf"], interactive=True, label="Upload PDF documents")
                with gr.Row():
                    slider_chunk_size = gr.Slider(minimum=10, maximum=1000, value=200, step=5, label="Chunk Size")
                    slider_chunk_overlap = gr.Slider(minimum=0, maximum=512, value=20, step=5, label="Chunk Overlap")
                with gr.Row():
                    db_btn = gr.Button("Create Embeddings")
                with gr.Row():
                    db_progress = gr.Textbox(value="Not initialized", show_label=False)
                gr.Markdown("<style>body { font-size: 16px; }</style><b>Select Large Language Model (LLM)</b>")
                with gr.Row():
                    llm_btn = gr.Radio(list_llm_simple, label="Available LLMs", value=list_llm_simple[0], type="index")
                with gr.Row():
                    qachain_btn = gr.Button("Initialize RAG system")
                with gr.Row():
                    llm_progress = gr.Textbox(value="Not initialized", show_label=False)

            with gr.Column(scale=200):
                gr.Markdown("<b>Step 2 - Chat with your Document</b>")
                chatbot = gr.Chatbot(height=505)
                with gr.Accordion("Similar context from the source document", open=False):
                    with gr.Row():
                        doc_source1 = gr.Textbox(label="Reference 1", lines=2, container=True, scale=20)
                        source1_page = gr.Number(label="Page", scale=1)
                    with gr.Row():
                        doc_source2 = gr.Textbox(label="Reference 2", lines=2, container=True, scale=20)
                        source2_page = gr.Number(label="Page", scale=1)
                    with gr.Row():
                        doc_source3 = gr.Textbox(label="Reference 3", lines=2, container=True, scale=20)
                        source3_page = gr.Number(label="Page", scale=1)
                with gr.Row():
                    msg = gr.Textbox(placeholder="Ask a question", container=True)
                with gr.Row():
                    submit_btn = gr.Button("Submit")
                    clear_btn = gr.ClearButton([msg, chatbot], value="Clear")

        # Preprocessing events
        db_btn.click(initialize_database, 
            inputs=[document, slider_chunk_size, slider_chunk_overlap], 
            outputs=[vector_db, db_progress])
        qachain_btn.click(initialize_LLM, 
            inputs=[llm_btn, vector_db], 
            outputs=[qa_chain, llm_progress]).then(lambda:[None, "", 0, "", 0, "", 0], 
            inputs=None, 
            outputs=[chatbot, doc_source1, source1_page, doc_source2, source2_page, doc_source3, source3_page], 
            queue=False)

        # Chatbot events
        msg.submit(conversation, 
            inputs=[qa_chain, msg, chatbot], 
            outputs=[qa_chain, msg, chatbot, doc_source1, source1_page, doc_source2, source2_page, doc_source3, source3_page], 
            queue=False)
        submit_btn.click(conversation, 
            inputs=[qa_chain, msg, chatbot], 
            outputs=[qa_chain, msg, chatbot, doc_source1, source1_page, doc_source2, source2_page, doc_source3, source3_page], 
            queue=False)
        clear_btn.click(lambda: [None, "", 0, "", 0, "", 0], 
            inputs=None, 
            outputs=[chatbot, doc_source1, source1_page, doc_source2, source2_page, doc_source3, source3_page], 
            queue=False)
    demo.queue().launch(debug=True)

if __name__ == "__main__":
    demo()