app.py

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

load_dotenv()

api_token = os.getenv("HF_TOKEN")

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]

# Load and split PDF document
def load_doc(list_file_path, chunk_size=512, chunk_overlap=64):
    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 with improved embedding model and parameters
def create_db(splits, n_trees=5, search_k=100):
    embeddings = HuggingFaceEmbeddings(model_name="sentence-transformers/all-mpnet-base-v2")
    vectordb = FAISS.from_documents(splits, embeddings, n_trees=n_trees, search_k=search_k)
    return vectordb

# Query expansion and document filtering functions
def expand_query(query):
    expanded_queries = [query, query + " additional term", query + " another term"]
    return expanded_queries

def filter_documents(docs):
    filtered_docs = [doc for doc in docs if "important" in doc.page_content]
    return filtered_docs

# Initialize langchain LLM chain with query expansion and document filtering
def initialize_llmchain(llm_model, temperature, max_tokens, top_k, vector_db, progress=gr.Progress()):
    llm = HuggingFaceEndpoint(
        repo_id=llm_model,
        huggingfacehub_api_token=api_token,
        temperature=temperature,
        max_new_tokens=max_tokens,
        top_k=top_k,
    )

    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,
        query_expansion=expand_query,
        document_filtering=filter_documents
    )
    return qa_chain

# Pre-process and vectorize local PDFs
def pre_process_pdfs(directory="pdfs"):
    file_paths = [os.path.join(directory, f) for f in os.listdir(directory) if f.endswith('.pdf')]
    doc_splits = load_doc(file_paths)
    vector_db = create_db(doc_splits)
    return vector_db

# 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 LLM
def initialize_LLM(llm_option, llm_temperature, max_tokens, top_k, vector_db, progress=gr.Progress()):
    llm_name = list_llm[llm_option]
    print("llm_name: ", llm_name)
    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

# Read persona from .md file
def load_persona(file_path):
    with open(file_path, 'r') as file:
        return file.read()

# Inject persona into response
def persona_template(response_text, persona_text):
    return f"{persona_text}\n\n{response_text}"

def conversation(qa_chain, message, history, persona_text):
    formatted_chat_history = format_chat_history(message, history)
    response = qa_chain.invoke({"question": message, "chat_history": formatted_chat_history})
    response_answer = response["answer"]
    if "Helpful Answer:" in response_answer:
        response_answer = response_answer.split("Helpful Answer:")[-1]
    response_answer = persona_template(response_answer, persona_text)
    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, response2_page, response_source3, source3_page

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

def demo():
    persona_text = load_persona('persona.md')

    with gr.Blocks(theme=gr.themes.Default(primary_hue="sky")) as demo:
        vector_db = gr.State(pre_process_pdfs("ILYA/pdfs"))  # Pre-process PDFs on initialization with correct path
        qa_chain = gr.State()
        gr.HTML("<center><h1>RAG PDF Chatbot</h1><center>")
        gr.Markdown("""<b>Interact with Your PDF Documents!</b> This AI agent performs retrieval-augmented generation (RAG) on PDF documents. Hosted on Hugging Face Hub for demonstration purposes. \
        <b>Do not upload confidential documents.</b>""")
        
        # Interface for static pre-selected documents
        gr.Markdown("<b>Pre-Selected Documents</b>")
        gr.Textbox(value="Document 1: Introduction to AI.pdf", show_label=False, interactive=False)
        gr.Textbox(value="Document 2: Advanced Machine Learning.pdf", show_label=False, interactive=False)
        
        gr.Markdown("<b>Upload Your PDF Documents</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 Configure Parameters</b>")
        llm_btn = gr.Radio(list_llm_simple, label="Available LLMs", value=list_llm_simple[0], type="index")
        slider_temperature = gr.Slider(minimum=0.01, maximum=1.0, value=0.5, step=0.1, label="Temperature", info="Controls randomness in token generation", interactive=True)
        slider_maxtokens = gr.Slider(minimum=128, maximum=9192, value=4096, step=128, label="Max New Tokens", info="Maximum number of tokens to be generated", interactive=True)
        slider_topk = gr.Slider(minimum=1, maximum=10, value=3, step=1, label="Top-K", info="Number of tokens to select the next token from", interactive=True)
        qachain_btn = gr.Button("Initialize Question Answering Chatbot")
        llm_progress = gr.Textbox(value="Not initialized", show_label=False)
        
        gr.Markdown("<b>Chat with your Document</b>")
        chatbot = gr.Chatbot(height=505)
        with gr.Accordion("Relevant 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)
        msg = gr.Textbox(placeholder="Ask a question", container=True)
        submit_btn = gr.Button("Submit")
        clear_btn = gr.ClearButton([msg, chatbot], value="Clear")

        # Preprocessing events
        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, gr.State(value=persona_text)], 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, gr.State(value=persona_text)], 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()