import gradio as gr # used for UI dev
import os # Built-in model to get/use the token for running huggingface source model which requires token to run
from typing import List, Dict # Built-in model 
from langchain.text_splitter import ( # Text splitting strategies 
    RecursiveCharacterTextSplitter,#Text splitting strategies  
    CharacterTextSplitter,#Text splitting strategies 
    TokenTextSplitter#Text splitting strategies 
)

from langchain_community.vectorstores import FAISS, Chroma, Qdrant # Vector database
from langchain_community.document_loaders import PyPDFLoader # Convert PDF to TEXT
from langchain.chains import ConversationalRetrievalChain # Entire retrival chain for conversation
from langchain_community.embeddings import HuggingFaceEmbeddings # Words to no
from langchain_huggingface import HuggingFaceEndpoint # API for generative model
from langchain.memory import ConversationBufferMemory # Chat History

list_llm = ["meta-llama/Meta-Llama-3-8B-Instruct", "mistralai/Mistral-7B-Instruct-v0.2"] # list of model
list_llm_simple = [os.path.basename(llm) for llm in list_llm] # display purpose
api_token = os.getenv("HF_TOKEN") # getting token
# Defining Chunk sizes
CHUNK_SIZES = {
    "small": {"recursive": 512, "fixed": 512, "token": 256},
    "medium": {"recursive": 1024, "fixed": 1024, "token": 512}
}
# passing Strategy , Chunk size , overlap
def get_text_splitter(strategy: str, chunk_size: int = 1024, chunk_overlap: int = 64):
    splitters = {
        "recursive": RecursiveCharacterTextSplitter(
            chunk_size=chunk_size,
            chunk_overlap=chunk_overlap
        ),
        "fixed": CharacterTextSplitter(
            chunk_size=chunk_size,
            chunk_overlap=chunk_overlap
        ),
        "token": TokenTextSplitter(
            chunk_size=chunk_size,
            chunk_overlap=chunk_overlap
        )
    }
    return splitters.get(strategy)

# def get_text_splitter(strategy, chunk_size=1024, chunk_overlap=64):
#     if strategy == "recursive":
#         return RecursiveCharacterTextSplitter(chunk_size=chunk_size, chunk_overlap=chunk_overlap)
#     elif strategy == "fixed":
#         return CharacterTextSplitter(chunk_size=chunk_size, chunk_overlap=chunk_overlap)
#     elif strategy == "token":
#         return TokenTextSplitter(chunk_size=chunk_size, chunk_overlap=chunk_overlap)
#     return None

def load_doc(list_file_path: List[str], splitting_strategy: str, chunk_size: str):
    chunk_size_value = CHUNK_SIZES[chunk_size][splitting_strategy]
    loaders = [PyPDFLoader(x) for x in list_file_path]
    pages = []
    for loader in loaders:
        pages.extend(loader.load())
    
    text_splitter = get_text_splitter(splitting_strategy, chunk_size_value)
    doc_splits = text_splitter.split_documents(pages)
    return doc_splits

def create_db(splits, db_choice: str = "faiss"):
    embeddings = HuggingFaceEmbeddings()
    db_creators = {
        "faiss": lambda: FAISS.from_documents(splits, embeddings),
        "chroma": lambda: Chroma.from_documents(splits, embeddings),
        "qdrant": lambda: Qdrant.from_documents(
            splits,
            embeddings,
            location=":memory:",   # In memory database for qdrant 
            collection_name="pdf_docs"
        )
    }
    return db_creators[db_choice]()

# Initialize Vector DB
def initialize_database(list_file_obj, splitting_strategy, chunk_size, db_choice, progress=gr.Progress()):
    """Initialize vector database with error handling"""
    try:
        if not list_file_obj:
            return None, "No files uploaded. Please upload PDF documents first."
            
        list_file_path = [x.name for x in list_file_obj if x is not None]
        if not list_file_path:
            return None, "No valid files found. Please upload PDF documents."
            
        doc_splits = load_doc(list_file_path, splitting_strategy, chunk_size)
        if not doc_splits:
            return None, "No content extracted from documents." 
            
        vector_db = create_db(doc_splits, db_choice)
        return vector_db, f"Database created successfully using {splitting_strategy} splitting and {db_choice} vector database!"
        
    except Exception as e:
        return None, f"Error creating database: {str(e)}"

def initialize_llmchain(llm_choice, temperature, max_tokens, top_k, vector_db, progress=gr.Progress()):
    """Initialize LLM chain with error handling"""
    try:
        if vector_db is None:
            return None, "Please create vector database first."
            
        llm_model = list_llm[llm_choice]
        
        llm = HuggingFaceEndpoint(
            repo_id=llm_model,
            huggingfacehub_api_token=api_token,
            temperature=temperature,
            max_new_tokens=max_tokens,
            top_k=top_k
        )
        # Temporary memory
        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,
            memory=memory,
            return_source_documents=True
        )
        return qa_chain, "LLM initialized successfully!"
        
    except Exception as e:
        return None, f"Error initializing LLM: {str(e)}"

def conversation(qa_chain, message, history):
    """Conversation function returning all required outputs"""
    response = qa_chain.invoke({
        "question": message,
        "chat_history": [(hist[0], hist[1]) for hist in history]
    })
    
    response_answer = response["answer"]
    if "Helpful Answer:" in response_answer:
        response_answer = response_answer.split("Helpful Answer:")[-1]
    
    sources = response["source_documents"][:3]
    source_contents = []
    source_pages = []
    
    for source in sources:
        source_contents.append(source.page_content.strip())
        source_pages.append(source.metadata.get("page", 0) + 1)
    
    while len(source_contents) < 3:
        source_contents.append("")
        source_pages.append(0)
    
    return (
        qa_chain,  
        gr.update(value=""), 
        history + [(message, response_answer)], 
        source_contents[0],  
        source_pages[0],   
        source_contents[1],  
        source_pages[1],    
        source_contents[2], 
        source_pages[2]     
    )

def demo():
    with gr.Blocks(theme=gr.themes.Default(primary_hue="red", secondary_hue="pink", neutral_hue="sky")) as demo:
        vector_db = gr.State()
        qa_chain = gr.State()
        
        gr.HTML("<center><h1>RAG PDF Chatbot</h1></center>")
        
        with gr.Column(scale=86):
            gr.Markdown("<b>Step 1 - Configure and Initialize RAG Pipeline</b>")
            with gr.Row():
                document = gr.Files(
                    height=300,
                    file_count="multiple",
                    file_types=["pdf"],
                    interactive=True,
                    label="Upload PDF documents"
                )
            
            with gr.Row():
                splitting_strategy = gr.Radio(
                    ["recursive", "fixed", "token"],
                    label="Text Splitting Strategy",
                    value="recursive"
                )
                db_choice = gr.Radio(
                    ["faiss", "chroma", "qdrant"],
                    label="Vector Database",
                    value="faiss"
                )
                chunk_size = gr.Radio(
                    ["small", "medium"],
                    label="Chunk Size",
                    value="medium"
                )
            
            with gr.Row():
                db_btn = gr.Button("Create vector database")
                db_progress = gr.Textbox(
                    value="Not initialized",
                    show_label=False
                )
            
            gr.Markdown("<b>Step 2 - Configure LLM</b>")
            with gr.Row():
                llm_choice = gr.Radio(
                    list_llm_simple,
                    label="Available LLMs",
                    value=list_llm_simple[0],
                    type="index"
                )
            
            with gr.Row():
                with gr.Accordion("LLM Parameters", open=False):
                    temperature = gr.Slider(
                        minimum=0.01,
                        maximum=1.0,
                        value=0.5,
                        step=0.1,
                        label="Temperature"
                    )
                    max_tokens = gr.Slider(
                        minimum=128,
                        maximum=4096,
                        value=2048,
                        step=128,
                        label="Max Tokens"
                    )
                    top_k = gr.Slider(
                        minimum=1,
                        maximum=10,
                        value=3,
                        step=1,
                        label="Top K"
                    )
            
            with gr.Row():
                init_llm_btn = gr.Button("Initialize LLM")
                llm_progress = gr.Textbox(
                    value="Not initialized",
                    show_label=False
                )
        
        with gr.Column(scale=200):
            gr.Markdown("<b>Step 3 - Chat with Documents</b>")
            chatbot = gr.Chatbot(height=505)
            
            with gr.Accordion("Source References", open=False):
                with gr.Row():
                    source1 = gr.Textbox(label="Source 1", lines=2)
                    page1 = gr.Number(label="Page")
                with gr.Row():
                    source2 = gr.Textbox(label="Source 2", lines=2)
                    page2 = gr.Number(label="Page")
                with gr.Row():
                    source3 = gr.Textbox(label="Source 3", lines=2)
                    page3 = gr.Number(label="Page")
            
            with gr.Row():
                msg = gr.Textbox(
                    placeholder="Ask a question",
                    show_label=False
                )
            with gr.Row():
                submit_btn = gr.Button("Submit")
                clear_btn = gr.ClearButton(
                    [msg, chatbot],
                    value="Clear Chat"
                )

        # Event handlers
        db_btn.click(
            initialize_database,
            inputs=[document, splitting_strategy, chunk_size, db_choice],
            outputs=[vector_db, db_progress]
        ).then(
            lambda x: gr.update(interactive=True) if x[0] is not None else gr.update(interactive=False),
            inputs=[vector_db],
            outputs=[init_llm_btn]
        )
        
        init_llm_btn.click(
            initialize_llmchain,
            inputs=[llm_choice, temperature, max_tokens, top_k, vector_db],
            outputs=[qa_chain, llm_progress]
        ).then(
            lambda x: gr.update(interactive=True) if x[0] is not None else gr.update(interactive=False),
            inputs=[qa_chain],
            outputs=[msg]
        )
        
        msg.submit(
            conversation,
            inputs=[qa_chain, msg, chatbot],
            outputs=[qa_chain, msg, chatbot, source1, page1, source2, page2, source3, page3]
        )
        
        submit_btn.click(
            conversation,
            inputs=[qa_chain, msg, chatbot],
            outputs=[qa_chain, msg, chatbot, source1, page1, source2, page2, source3, page3]
        )

        clear_btn.click(
            lambda: [None, "", 0, "", 0, "", 0],
            outputs=[chatbot, source1, page1, source2, page2, source3, page3]
        )

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

if __name__ == "__main__":
    demo()