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]

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 
    )  
    return text_splitter.split_documents(pages)

def create_db(splits):
    return FAISS.from_documents(splits, HuggingFaceEmbeddings())

def initialize_llmchain(llm_model, temperature, max_tokens, top_k, vector_db):
    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
    )

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

def initialize_database(list_file_obj):
    list_file_path = [x.name for x in list_file_obj if x is not None]
    return create_db(load_doc(list_file_path)), "Database created!"

def initialize_LLM(llm_option, llm_temperature, max_tokens, top_k, vector_db):
    llm_name = list_llm[llm_option]
    return initialize_llmchain(llm_name, llm_temperature, max_tokens, top_k, vector_db), "QA chain initialized!"

def format_chat_history(message, chat_history):
    formatted = []
    for user_msg, bot_msg in chat_history:
        formatted.extend([f"User: {user_msg}", f"Assistant: {bot_msg}"])
    return formatted

def conversation(qa_chain, message, history):
    response = qa_chain.invoke({
        "question": message, 
        "chat_history": format_chat_history(message, history)
    })
    answer = response["answer"].split("Helpful Answer:")[-1]
    sources = response["source_documents"]
    return (
        qa_chain, 
        gr.update(value=""), 
        history + [(message, answer)],
        sources[0].page_content.strip(), sources[0].metadata["page"] + 1,
        sources[1].page_content.strip(), sources[1].metadata["page"] + 1,
        sources[2].page_content.strip(), sources[2].metadata["page"] + 1
    )

def demo():
    css = """
    #main-container {
        display: flex;
        flex-wrap: nowrap;
        width: 80% !important;
        max-width: 100vw !important;
        overflow: hidden;
    }
    #left-column {
        flex: 0 1 35%;
        min-width: 150px !important;
        max-width: 100% !important;
    }
    #right-column {
        flex: 1 1 65%;
        min-width: 300px !important;
        max-width: 100% !important;
    }
    .chatbot {
        width: 100% !important;
        max-width: 100% !important;
    }
    .textbox {
        max-width: 100% !important;
    }
    @media (max-width: 900px) {
        #main-container { flex-wrap: wrap; }
        #left-column, #right-column { flex: 1 1 100% !important; }
    }
    """

    with gr.Blocks(theme=gr.themes.Default(primary_hue="red", secondary_hue="pink", neutral_hue="sky"), css=css) as demo:
        vector_db = gr.State()
        qa_chain = gr.State()
        
        gr.Markdown("## RAG PDF Chatbot")
        
        with gr.Row(elem_id="main-container"):
            with gr.Column(elem_id="left-column"):
                # Step 1 Column
                with gr.Group():
                    gr.Markdown("**Step 1 - Setup**")
                    docs = gr.Files(file_types=[".pdf"], label="Upload PDFs")
                    db_btn = gr.Button("Create Vector DB")
                    db_status = gr.Textbox("Not initialized", show_label=False)
                    
                    gr.Markdown("**LLM Selection**")
                    llm_select = gr.Radio(list_llm_simple, label="Model", value=list_llm_simple[0], type="index")
                    
                    with gr.Accordion("Parameters", open=False):
                        temp = gr.Slider(0.01, 1.0, 0.5, label="Temperature")
                        tokens = gr.Slider(128, 8192, 4096, step=128, label="Max Tokens")
                        topk = gr.Slider(1, 10, 3, step=1, label="Top-K")
                    
                    init_btn = gr.Button("Initialize Chatbot")
                    llm_status = gr.Textbox("Not initialized", show_label=False)

            with gr.Column(elem_id="right-column"):
                # Step 2 Column
                gr.Markdown("**Step 2 - Chat**")
                chatbot = gr.Chatbot(height=400, elem_classes="chatbot")
                
                with gr.Accordion("Source Context", open=False):
                    with gr.Row():
                        src1 = gr.Textbox(label="Reference 1", lines=2, max_lines=2, elem_classes="textbox")
                        pg1 = gr.Number(label="Page")
                    with gr.Row():
                        src2 = gr.Textbox(label="Reference 2", lines=2, max_lines=2, elem_classes="textbox")
                        pg2 = gr.Number(label="Page")
                    with gr.Row():
                        src3 = gr.Textbox(label="Reference 3", lines=2, max_lines=2, elem_classes="textbox")
                        pg3 = gr.Number(label="Page")
                
                msg = gr.Textbox(placeholder="Ask something...", elem_classes="textbox")
                with gr.Row():
                    submit = gr.Button("Submit")
                    clear = gr.ClearButton([msg, chatbot])

        # Event handlers
        db_btn.click(
            initialize_database, [docs], [vector_db, db_status]
        )
        init_btn.click(
            initialize_LLM, [llm_select, temp, tokens, topk, vector_db], [qa_chain, llm_status]
        ).then(
            lambda: [None,"",0,"",0,"",0], None, [chatbot, src1, pg1, src2, pg2, src3, pg3]
        )
        msg.submit(conversation, [qa_chain, msg, chatbot], 
                  [qa_chain, msg, chatbot, src1, pg1, src2, pg2, src3, pg3])
        submit.click(conversation, [qa_chain, msg, chatbot], 
                   [qa_chain, msg, chatbot, src1, pg1, src2, pg2, src3, pg3])
        clear.click(
            lambda: [None,"",0,"",0,"",0], None, [chatbot, src1, pg1, src2, pg2, src3, pg3]
        )

    return demo

if __name__ == "__main__":
    demo().launch()