app.py

import gradio as gr
import os
api_token = os.getenv("HF_TOKEN")

from langchain_community.vectorstores import FAISS
from langchain_community.document_loaders import PyPDFLoader
from langchain.text_splitter import RecursiveCharacterTextSplitter
from langchain.chains import ConversationalRetrievalChain
from langchain_community.embeddings import HuggingFaceEmbeddings 
from langchain.memory import ConversationBufferMemory
from langchain_community.llms import HuggingFaceEndpoint

list_llm = ["meta-llama/Meta-Llama-3-8B-Instruct", "mistralai/Mistral-7B-Instruct-v0.2"]  
list_llm_simple = [os.path.basename(llm) for llm in list_llm]

# Simulated user credentials (replace with a secure method in production)
VALID_USERNAME = "admin"
VALID_PASSWORD = "password123"

# Load and split PDF document
def load_doc(list_file_path):
    loaders = [PyPDFLoader(x) for x in list_file_path]
    pages = []
    for loader in loaders:
        pages.extend(loader.load())
    text_splitter = RecursiveCharacterTextSplitter(chunk_size=1024, chunk_overlap=64)
    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 database
def initialize_database(list_file_obj, 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)
    vector_db = create_db(doc_splits)
    return vector_db, "Database created!"

# Initialize langchain LLM chain
def initialize_llmchain(llm_model, temperature, max_tokens, top_k, vector_db, progress=gr.Progress()):
    if llm_model == "meta-llama/Meta-Llama-3-8B-Instruct":
        llm = HuggingFaceEndpoint(
            repo_id=llm_model,
            huggingfacehub_api_token=api_token,
            temperature=temperature,
            max_new_tokens=max_tokens,
            top_k=top_k,
            timeout=120,
            max_retries=3
        )
    else:
        llm = HuggingFaceEndpoint(
            huggingfacehub_api_token=api_token,
            repo_id=llm_model, 
            temperature=temperature,
            max_new_tokens=max_tokens,
            top_k=top_k,
            timeout=120,
            max_retries=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 LLM
def initialize_LLM(llm_option, llm_temperature, max_tokens, top_k, vector_db, progress=gr.Progress()):
    llm_name = list_llm[llm_option]
    qa_chain = initialize_llmchain(llm_name, llm_temperature, max_tokens, top_k, vector_db, progress)
    return qa_chain, "QA chain initialized. Chatbot is ready!"

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, language):
    formatted_chat_history = format_chat_history(message, history)
    if language == "Português":
        prompt = f"Responda em português: {message}"
    else:
        prompt = f"Answer in English: {message}"
    
    try:
        response = qa_chain.invoke({"question": prompt, "chat_history": formatted_chat_history})
        response_answer = response["answer"]
        if response_answer.find("Helpful Answer:") != -1:
            response_answer = response_answer.split("Helpful Answer:")[-1]
    except Exception as e:
        if language == "Português":
            response_answer = f"Erro: Não foi possível obter resposta do modelo devido a problemas no servidor. Tente novamente mais tarde. ({str(e)})"
        else:
            response_answer = f"Error: Could not get a response from the model due to server issues. Please try again later. ({str(e)})"
    
    try:
        response_sources = response["source_documents"]
        response_source1 = response_sources[0].page_content.strip()
        response_source1_page = response_sources[0].metadata["page"] + 1
        response_source2 = response_sources[1].page_content.strip()
        response_source2_page = response_sources[1].metadata["page"] + 1
        response_source3 = response_sources[2].page_content.strip()
        response_source3_page = response_sources[2].metadata["page"] + 1
    except:
        response_source1 = response_source2 = response_source3 = "N/A"
        response_source1_page = response_source2_page = response_source3_page = 0
    
    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

# Login function
def check_login(username, password):
    if username == VALID_USERNAME and password == VALID_PASSWORD:
        return True, "Login successful! Access the chatbot below."
    else:
        return False, "Invalid username or password. Please try again."

# Main demo with login
def demo():
    with gr.Blocks(theme=gr.themes.Default(primary_hue="red", secondary_hue="pink", neutral_hue="sky")) as demo:
        # State variables
        vector_db = gr.State()
        qa_chain = gr.State()
        logged_in = gr.State(value=False)

        # Login interface
        with gr.Column(visible=True) as login_col:
            gr.HTML("<center><h1>RAG PDF Chatbot - Login</h1></center>")
            username = gr.Textbox(label="Username", placeholder="Enter username")
            password = gr.Textbox(label="Password", type="password", placeholder="Enter password")
            login_btn = gr.Button("Login")
            login_message = gr.Textbox(value="Please log in to access the chatbot.", show_label=False)

        # Chatbot interface (hidden until login)
        with gr.Column(visible=False) as chatbot_col:
            gr.HTML("<center><h1>RAG PDF Chatbot</h1></center>")
            gr.Markdown("""<b>Query your PDF documents!</b> This AI agent 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 RAG pipeline</b>")
                    document = gr.Files(height=300, file_count="multiple", file_types=["pdf"], interactive=True, label="Upload PDF documents")
                    db_btn = gr.Button("Create vector database")
                    db_progress = gr.Textbox(value="Not initialized", show_label=False)
                    gr.Markdown("<b>Select Large Language Model (LLM) and input parameters</b>")
                    llm_btn = gr.Radio(list_llm_simple, label="Available LLMs", value=list_llm_simple[0], type="index")
                    with gr.Accordion("LLM input parameters", open=False):
                        slider_temperature = gr.Slider(minimum=0.01, maximum=1.0, value=0.5, step=0.1, label="Temperature", interactive=True)
                        slider_maxtokens = gr.Slider(minimum=128, maximum=9192, value=4096, step=128, label="Max New Tokens", interactive=True)
                        slider_topk = gr.Slider(minimum=1, maximum=10, value=3, step=1, label="top-k", interactive=True)
                    qachain_btn = gr.Button("Initialize Question Answering Chatbot")
                    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>")
                    language_selector = gr.Radio(["English", "Português"], label="Select Language", value="English")
                    chatbot = gr.Chatbot(height=505)
                    with gr.Accordion("Relevant context from the source document", open=False):
                        doc_source1 = gr.Textbox(label="Reference 1", lines=2, container=True, scale=20)
                        source1_page = gr.Number(label="Page", scale=1)
                        doc_source2 = gr.Textbox(label="Reference 2", lines=2, container=True, scale=20)
                        source2_page = gr.Number(label="Page", scale=1)
                        doc_source3 = gr.Textbox(label="Reference 3", lines=2, container=True, scale=20)
                        source3_page = gr.Number(label="Page", scale=1)
                    msg = gr.Textbox(placeholder="Ask a question", container=True)
                    submit_btn = gr.Button("Submit")
                    clear_btn = gr.ClearButton([msg, chatbot], value="Clear")

        # Login event
        login_btn.click(
            fn=check_login,
            inputs=[username, password],
            outputs=[logged_in, login_message]
        ).then(
            fn=lambda logged: (gr.update(visible=not logged), gr.update(visible=logged)),
            inputs=[logged_in],
            outputs=[login_col, chatbot_col],
            queue=False
        )

        # Preprocessing events (only accessible after login)
        db_btn.click(initialize_database, inputs=[document], outputs=[vector_db, db_progress])
        qachain_btn.click(initialize_LLM, inputs=[llm_btn, slider_temperature, slider_maxtokens, slider_topk, 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, language_selector], 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, language_selector], 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()