app.py

import gradio as gr
import os
from pathlib import Path
import re
from unidecode import unidecode
import chromadb
from langchain_community.vectorstores import FAISS, ScaNN, Milvus
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
from huggingface_hub import InferenceClient
import torch

api_token = os.getenv("HF_TOKEN")

client = InferenceClient(
    "mistralai/Mistral-7B-Instruct-v0.3"
)

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

def format_prompt(message, history):
    prompt = "<s>"
    for user_prompt, bot_response in history:
        prompt += f"[INST] {user_prompt} [/INST]"
        prompt += f" {bot_response}</s> "
    prompt += f"[INST] {message} [/INST]"
    return prompt

def generate(prompt, history, temperature=0.9, max_new_tokens=256, top_p=0.95, repetition_penalty=1.0):
    temperature = float(temperature)
    if temperature < 1e-2:
        temperature = 1e-2
    top_p = float(top_p)

    generate_kwargs = dict(
        temperature=temperature,
        max_new_tokens=max_new_tokens,
        top_p=top_p,
        repetition_penalty=repetition_penalty,
        do_sample=True,
        seed=42,
    )

    formatted_prompt = format_prompt(prompt, history)

    stream = client.text_generation(formatted_prompt, **generate_kwargs, stream=True, details=True, return_full_text=False)
    output = ""

    for response in stream:
        output += response.token.text
        yield output
    return output

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

def create_db(splits, collection_name, db_type):
    embedding = HuggingFaceEmbeddings()
    
    if db_type == "ChromaDB":
        new_client = chromadb.EphemeralClient()
        vectordb = Chroma.from_documents(
            documents=splits,
            embedding=embedding,
            client=new_client,
            collection_name=collection_name,
        )
    elif db_type == "FAISS":
        vectordb = FAISS.from_documents(
            documents=splits,
            embedding=embedding
        )
    elif db_type == "ScaNN":
        vectordb = ScaNN.from_documents(
            documents=splits,
            embedding=embedding
        )
    elif db_type == "Milvus":
        vectordb = Milvus.from_documents(
            documents=splits,
            embedding=embedding,
            collection_name=collection_name,
        )
    else:
        raise ValueError(f"Unsupported vector database type: {db_type}")
    
    return vectordb

def initialize_llmchain(llm_model, temperature, max_tokens, top_k, vector_db, initial_prompt, progress=gr.Progress()):
    progress(0.1, desc="Initializing HF tokenizer...")
    progress(0.5, desc="Initializing HF Hub...")

    llm = HuggingFaceEndpoint(
        repo_id=llm_model,
        huggingfacehub_api_token=api_token,
        temperature=temperature,
        max_new_tokens=max_tokens,
        top_k=top_k,
    )

    progress(0.75, desc="Defining buffer memory...")
    memory = ConversationBufferMemory(
        memory_key="chat_history",
        output_key='answer',
        return_messages=True
    )
    retriever = vector_db.as_retriever()
    progress(0.8, desc="Defining retrieval chain...")
    qa_chain = ConversationalRetrievalChain.from_llm(
        llm,
        retriever=retriever,
        chain_type="stuff",
        memory=memory,
        return_source_documents=True,
        verbose=False,
    )
    qa_chain({"question": initial_prompt})  # Initialize with the initial prompt
    progress(0.9, desc="Done!")
    return qa_chain

def initialize_llm_no_doc(llm_model, temperature, max_tokens, top_k, initial_prompt, progress=gr.Progress()):
    progress(0.1, desc="Initializing HF tokenizer...")
    progress(0.5, desc="Initializing HF Hub...")
    llm = HuggingFaceEndpoint(
        repo_id=llm_model,
        huggingfacehub_api_token=api_token,
        temperature=temperature,
        max_new_tokens=max_tokens,
        top_k=top_k,
    )
    progress(0.75, desc="Defining buffer memory...")
    memory = ConversationBufferMemory(
        memory_key="chat_history",
        output_key='answer',
        return_messages=True
    )
    conversation_chain = ConversationChain(llm=llm, memory=memory, verbose=False)
    conversation_chain({"question": initial_prompt})
    progress(0.9, desc="Done!")
    return conversation_chain

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({"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_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 conversation_no_doc(llm, message, history):
    formatted_chat_history = format_chat_history(message, history)
    response = llm({"question": message, "chat_history": formatted_chat_history})
    response_answer = response["answer"]
    new_history = history + [(message, response_answer)]
    return llm, gr.update(value=""), new_history

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

def demo():
    with gr.Blocks(theme="base") as demo:
        vector_db = gr.State()
        qa_chain = gr.State()
        collection_name = gr.State()
        initial_prompt = gr.State("")
        llm_no_doc = gr.State()

        gr.Markdown(
        """<center><h2>lucIAna</center></h2>
        <h3>Olá, sou a 2. versão</h3>""")
        gr.Markdown(
        """<b>Note:</b> Esta é a lucIAna, primeira Versão da IA para seus PDF documentos. 
        Este chatbot leva em consideração perguntas anteriores ao gerar respostas (por meio de memória conversacional) e inclui referências a documentos para fins de clareza.
        """)

        with gr.Tab("Step 1 - Upload PDF"):
            with gr.Row():
                document = gr.Files(height=100, file_count="multiple", file_types=["pdf"], interactive=True, label="Upload your PDF documents (single or multiple)")
        
        with gr.Tab("Step 2 - Process document"):
            with gr.Row():
                db_type_radio = gr.Radio(["ChromaDB", "FAISS", "ScaNN", "Milvus"], label="Vector database type", value="ChromaDB", type="value", info="Choose your vector database")
            with gr.Accordion("Advanced options - Document text splitter", open=False):
                with gr.Row():
                    slider_chunk_size = gr.Slider(minimum=100, maximum=1000, value=600, step=20, label="Chunk size", info="Chunk size", interactive=True)
                with gr.Row():
                    slider_chunk_overlap = gr.Slider(minimum=10, maximum=200, value=40, step=10, label="Chunk overlap", info="Chunk overlap", interactive=True)
            with gr.Row():
                db_progress = gr.Textbox(label="Vector database initialization", value="None")
            with gr.Row():
                db_btn = gr.Button("Generate vector database")
        
        with gr.Tab("Step 3 - Set Initial Prompt"):
            with gr.Row():
                prompt_input = gr.Textbox(label="Initial Prompt", lines=5, value="Você é um advogado sênior, onde seu papel é analisar e trazer as informações sem inventar, dando a sua melhor opinião sempre trazendo contexto e referência. Aprenda o que é jurisprudência.")
            with gr.Row():
                set_prompt_btn = gr.Button("Set Prompt")

        with gr.Tab("Step 4 - Initialize QA chain"):
            with gr.Row():
                llm_btn = gr.Radio(list_llm_simple, 
                    label="LLM models", value=list_llm_simple[0], type="index", info="Choose your LLM model")
            with gr.Accordion("Advanced options - LLM model", open=False):
                with gr.Row():
                    slider_temperature = gr.Slider(minimum=0.01, maximum=1.0, value=0.7, step=0.1, label="Temperature", info="Model temperature", interactive=True)
                with gr.Row():
                    slider_maxtokens = gr.Slider(minimum=224, maximum=4096, value=1024, step=32, label="Max Tokens", info="Model max tokens", interactive=True)
                with gr.Row():
                    slider_topk = gr.Slider(minimum=1, maximum=10, value=3, step=1, label="top-k samples", info="Model top-k samples", interactive=True)
            with gr.Row():
                llm_progress = gr.Textbox(value="None", label="QA chain initialization")
            with gr.Row():
                qachain_btn = gr.Button("Initialize Question Answering chain")

        with gr.Tab("Step 5 - Chatbot with document"):
            chatbot = gr.Chatbot(height=300)
            with gr.Accordion("Advanced - Document references", 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="Type message (e.g. 'What is this document about?')", container=True)
            with gr.Row():
                submit_btn = gr.Button("Submit message")
                clear_btn = gr.ClearButton([msg, chatbot], value="Clear conversation")

        with gr.Tab("Step 6 - Chatbot without document"):
            chatbot_no_doc = gr.Chatbot(height=300)
            additional_inputs=[
                gr.Slider(
                    label="Temperature",
                    value=0.9,
                    minimum=0.0,
                    maximum=1.0,
                    step=0.05,
                    interactive=True,
                    info="Higher values produce more diverse outputs",
                ),
                gr.Slider(
                    label="Max new tokens",
                    value=256,
                    minimum=0,
                    maximum=1048,
                    step=64,
                    interactive=True,
                    info="The maximum numbers of new tokens",
                ),
                gr.Slider(
                    label="Top-p (nucleus sampling)",
                    value=0.90,
                    minimum=0.0,
                    maximum=1,
                    step=0.05,
                    interactive=True,
                    info="Higher values sample more low-probability tokens",
                ),
                gr.Slider(
                    label="Repetition penalty",
                    value=1.2,
                    minimum=1.0,
                    maximum=2.0,
                    step=0.05,
                    interactive=True,
                    info="Penalize repeated tokens",
                )
            ]
            with gr.Row():
                msg_no_doc = gr.Textbox(placeholder="Type message to chat with lucIAna", container=True)
            with gr.Row():
                submit_btn_no_doc = gr.Button("Submit message")
                clear_btn_no_doc = gr.ClearButton([msg_no_doc, chatbot_no_doc], value="Clear conversation")

            gr.ChatInterface(
                fn=generate,
                chatbot=chatbot_no_doc,
                additional_inputs=additional_inputs,
                title="Mistral 7B v0.3"
            )

        # Preprocessing events
        db_btn.click(initialize_database, 
            inputs=[document, slider_chunk_size, slider_chunk_overlap, db_type_radio], 
            outputs=[vector_db, collection_name, db_progress])
        set_prompt_btn.click(lambda prompt: gr.update(value=prompt), 
                             inputs=prompt_input, 
                             outputs=initial_prompt)
        qachain_btn.click(initialize_llmchain, 
            inputs=[llm_btn, slider_temperature, slider_maxtokens, slider_topk, vector_db, initial_prompt], 
            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 with document
        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)

        # Initialize LLM without document for conversation
        submit_btn_no_doc.click(conversation_no_doc, 
            inputs=[llm_no_doc, msg_no_doc, chatbot_no_doc], 
            outputs=[llm_no_doc, msg_no_doc, chatbot_no_doc], 
            queue=False)
        clear_btn_no_doc.click(lambda:[None,""], 
            inputs=None, 
            outputs=[chatbot_no_doc, msg_no_doc], 
            queue=False)

    demo.queue().launch(debug=True, share=True)

if __name__ == "__main__":
    demo()