app.py

import gradio as gr
import os

from langchain.document_loaders import PyPDFLoader
from langchain.text_splitter import RecursiveCharacterTextSplitter
from langchain.vectorstores import Chroma
from langchain.chains import ConversationalRetrievalChain
from langchain.embeddings import HuggingFaceEmbeddings 
from langchain.llms import HuggingFacePipeline
from langchain.chains import ConversationChain
from langchain.memory import ConversationBufferMemory
from langchain.llms import HuggingFaceHub

from pathlib import Path
import chromadb

from transformers import AutoTokenizer
import transformers
import torch
import tqdm 
import accelerate

from transformers import MBartForConditionalGeneration, MBart50TokenizerFast
translation_model = MBartForConditionalGeneration.from_pretrained("facebook/mbart-large-50-many-to-many-mmt")
translation_tokenizer = MBart50TokenizerFast.from_pretrained("facebook/mbart-large-50-many-to-many-mmt")

def english_to_hindi(sentence):
    translation_tokenizer.src_lang = "en_xx"
    encoded_hi = translation_tokenizer(sentence, return_tensors="pt")
    generated_tokens = translation_model.generate(**encoded_hi, forced_bos_token_id=translation_tokenizer.lang_code_to_id["hi_IN"] )
    return (translation_tokenizer.batch_decode(generated_tokens, skip_special_tokens=True))


def hindi_to_english(sentence):
    translation_tokenizer.src_lang = "hi_IN"
    encoded_hi = translation_tokenizer(sentence, return_tensors="pt")
    generated_tokens = translation_model.generate(**encoded_hi, forced_bos_token_id=translation_tokenizer.lang_code_to_id["en_XX"] )
    return (translation_tokenizer.batch_decode(generated_tokens, skip_special_tokens=True))


llm_model = "mistralai/Mistral-7B-Instruct-v0.2"


# default_persist_directory = './chroma_HF/'
list_llm = ["mistralai/Mistral-7B-Instruct-v0.2", "mistralai/Mixtral-8x7B-Instruct-v0.1", "mistralai/Mistral-7B-Instruct-v0.1", \
    "google/gemma-7b-it","google/gemma-2b-it", \
    "HuggingFaceH4/zephyr-7b-beta", "meta-llama/Llama-2-7b-chat-hf", "microsoft/phi-2", \
    "TinyLlama/TinyLlama-1.1B-Chat-v1.0", "mosaicml/mpt-7b-instruct", "tiiuae/falcon-7b-instruct", \
    "google/flan-t5-xxl"
]
list_llm_simple = [os.path.basename(llm) for llm in list_llm]

# Load PDF document and create doc splits
def load_doc(list_file_path, chunk_size, chunk_overlap):
    # Processing for one document only
    # loader = PyPDFLoader(file_path)
    # pages = loader.load()
    loaders = [PyPDFLoader(x) for x in list_file_path]
    pages = []
    for loader in loaders:
        pages.extend(loader.load())
    # text_splitter = RecursiveCharacterTextSplitter(chunk_size = 600, chunk_overlap = 50)
    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, collection_name):
    embedding = HuggingFaceEmbeddings()
    new_client = chromadb.EphemeralClient()
    vectordb = Chroma.from_documents(
        documents=splits,
        embedding=embedding,
        client=new_client,
        collection_name=collection_name,
        # persist_directory=default_persist_directory
    )
    return vectordb


# Load vector database
def load_db():
    embedding = HuggingFaceEmbeddings()
    vectordb = Chroma(
        # persist_directory=default_persist_directory, 
        embedding_function=embedding)
    return vectordb


# Initialize langchain LLM chain
def initialize_llmchain(temperature, max_tokens, top_k, vector_db, progress=gr.Progress()):
    progress(0.1, desc="Initializing HF tokenizer...")
    
    # HuggingFaceHub uses HF inference endpoints
    progress(0.5, desc="Initializing HF Hub...")
    # Use of trust_remote_code as model_kwargs
    # Warning: langchain issue
    # URL: https://github.com/langchain-ai/langchain/issues/6080

    llm = HuggingFaceHub(repo_id=llm_model, model_kwargs={"temperature": temperature, 
                                                          "max_new_tokens": max_tokens, 
                                                          "top_k": top_k, 
                                                          "load_in_8bit": True})
    
    progress(0.75, desc="Defining buffer memory...")
    memory = ConversationBufferMemory(
        memory_key="chat_history",
        output_key='answer',
        return_messages=True
    )
    # retriever=vector_db.as_retriever(search_type="similarity", search_kwargs={'k': 3})
    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,
        # combine_docs_chain_kwargs={"prompt": your_prompt})
        return_source_documents=True,
        #return_generated_question=False,
        verbose=False,
    )
    progress(0.9, desc="Done!")
    return qa_chain


# Initialize database
def initialize_database(list_file_obj, chunk_size, chunk_overlap, progress=gr.Progress()):
    # Create list of documents (when valid)
    list_file_path = [x.name for x in list_file_obj if x is not None]
    # Create collection_name for vector database
    progress(0.1, desc="Creating collection name...")
    collection_name = Path(list_file_path[0]).stem
    # Fix potential issues from naming convention
    ## Remove space
    collection_name = collection_name.replace(" ","-") 
    ## Limit lenght to 50 characters
    collection_name = collection_name[:50]
    ## Enforce start and end as alphanumeric character
    if not collection_name[0].isalnum():
        collection_name[0] = 'A'
    if not collection_name[-1].isalnum():
        collection_name[-1] = 'Z'
    # print('list_file_path: ', list_file_path)
    print('Collection name: ', collection_name)
    progress(0.25, desc="Loading document...")
    # Load document and create splits
    doc_splits = load_doc(list_file_path, chunk_size, chunk_overlap)
    # Create or load vector database
    progress(0.5, desc="Generating vector database...")
    # global vector_db
    vector_db = create_db(doc_splits, collection_name)
    progress(0.9, desc="Done!")
    return vector_db, collection_name, "Complete!"


def initialize_LLM(llm_temperature, max_tokens, top_k, vector_db, progress=gr.Progress()):
    # print("llm_option",llm_option)
    llm_name = llm_model
    print("llm_name: ",llm_name)
    qa_chain = initialize_llmchain(llm_temperature, max_tokens, top_k, vector_db, progress)
    return qa_chain, "Complete!"


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)
    #print("formatted_chat_history",formatted_chat_history)
   
    # Generate response using QA chain
    response = qa_chain({"question": message, "chat_history": formatted_chat_history})
    #response_answer = response["answer"]
    response_answer = english_to_hindi(response["answer"])[0]
    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()
    # Langchain sources are zero-based
    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
    # print ('chat response: ', response_answer)
    # print('DB source', response_sources)
    
    # Append user message and response to chat history
    new_history = history + [(message, response_answer)]
    # return gr.update(value=""), new_history, response_sources[0], response_sources[1] 
    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)
    # print(file_path)
    # initialize_database(file_path, progress)
    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()
        pdf_directory = '/home/user/app/pdfs'

        def process_pdfs():
            # List all PDF files in the directory
            pdf_files = [os.path.join(pdf_directory, file) for file in os.listdir(pdf_directory) if file.endswith(".pdf")]
            return pdf_files
            
        # Create a dictionary with the necessary information
        pdf_dict = {"value": process_pdfs, "height": 100, "file_count": "multiple", 
                    "visible": False, "file_types": ["pdf"], "interactive": True, 
                    "label": "Uploaded PDF documents"}
        
        # Create a gr.Files component with the dictionary
        #document_files = gr.Files(**pdf_dict)

        with gr.Row():
            # document = gr.Files(value = process_pdfs, height=100, file_count="multiple",visible=True,
            #                     file_types=["pdf"], interactive=True, label="Upload your PDF documents (single or multiple)")
            document = gr.Files(**pdf_dict)
        with gr.Row():
            db_btn = gr.Radio(["ChromaDB"], label="Vector database type", value = "ChromaDB", type="index", info="Choose your vector database",visible=False)
        with gr.Accordion("Advanced options - Document text splitter", open=False, visible=False):
            with gr.Row():
                slider_chunk_size = gr.Slider(value=2048, label="Chunk size", info="Chunk size", interactive=False, visible=False)
            with gr.Row():
                slider_chunk_overlap = gr.Slider(value=256, label="Chunk overlap", info="Chunk overlap", interactive=False, visible=False)

            


        with gr.Accordion("Advanced options - LLM model", open=False, visible=False):
            with gr.Row():
                slider_temperature = gr.Slider(value = 0.1,visible=False)
            with gr.Row():
                slider_maxtokens = gr.Slider(value = 4096, visible=False)
            with gr.Row():
                slider_topk = gr.Slider(value = 3, visible=False)



        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", container=True)
        with gr.Row():
            db_progress = gr.Textbox(label="Vector database initialization", value="None", visible=True)
        with gr.Row():
            db_btn = gr.Button("Generate vector database...")
        with gr.Row():
            llm_progress = gr.Textbox(value="None",label="QA chain initialization", visible=True)
        with gr.Row():
            qachain_btn = gr.Button("Initialize question-answering chain...")
        with gr.Row():
            submit_btn = gr.Button("Submit")
            clear_btn = gr.ClearButton([msg, chatbot])
            
        # Preprocessing events
        #upload_btn.upload(upload_file, inputs=[upload_btn], outputs=[document])
        db_btn.click(initialize_database, \
            inputs=[document, slider_chunk_size, slider_chunk_overlap], \
            outputs=[vector_db, collection_name, db_progress])
        qachain_btn.click(initialize_LLM, \
            inputs=[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], \
            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()