app.py

# app.py (Versi Final untuk Gradio di Hugging Face)

import gradio as gr
import os
import re
import shutil
from langchain_community.document_loaders import PyPDFLoader
from langchain.text_splitter import RecursiveCharacterTextSplitter
from langchain_community.embeddings import HuggingFaceEmbeddings
from langchain_community.vectorstores import FAISS
from langchain_community.retrievers import BM25Retriever
from langchain.retrievers import EnsembleRetriever
from transformers import AutoTokenizer, AutoModelForCausalLM
import torch

# --- 1. SETUP MODEL (dijalankan sekali saat aplikasi start) ---
@torch.no_grad()
def load_models():
    print("Memuat model (hanya terjadi sekali)...")
    device = "cuda" if torch.cuda.is_available() else "cpu"
    
    cache_dir = "./model_cache"
    os.makedirs(cache_dir, exist_ok=True)
    os.environ['SENTENCE_TRANSFORMERS_HOME'] = cache_dir

    embeddings = HuggingFaceEmbeddings(
        model_name="sentence-transformers/all-MiniLM-L6-v2",
        cache_folder=cache_dir
    )
    
    # Gunakan token dari secrets jika ada
    hf_token = os.getenv("HUGGING_FACE_HUB_TOKEN")

    tokenizer = AutoTokenizer.from_pretrained("google/gemma-3-270m-it", cache_dir=cache_dir, token=hf_token)
    llm = AutoModelForCausalLM.from_pretrained(
        "google/gemma-3-270m-it",
        cache_dir=cache_dir,
        device_map="auto",
        torch_dtype=torch.bfloat16,
        token=hf_token
    )
    print("Model berhasil dimuat.")
    return embeddings, tokenizer, llm

embeddings, tokenizer, llm = load_models()
# Inisialisasi state global untuk retriever dan chunks
rag_pipeline = {"retriever": None, "all_chunks": None}


# --- 2. FUNGSI INTI RAG (backend logic) ---
def process_document(uploaded_file):
    if uploaded_file is None:
        return "Mohon unggah file terlebih dahulu.", gr.update(interactive=False)
        
    try:
        # Gradio menyimpan file di temporary path, kita bisa langsung pakai
        file_path = uploaded_file.name
        
        loader = PyPDFLoader(file_path)
        docs = loader.load()
        text_splitter = RecursiveCharacterTextSplitter(chunk_size=2000, chunk_overlap=200)
        chunks = text_splitter.split_documents(docs)
        
        rag_pipeline["all_chunks"] = chunks
        
        faiss_db = FAISS.from_documents(chunks, embeddings)
        faiss_retriever = faiss_db.as_retriever(search_kwargs={"k": 10})
        bm25_retriever = BM25Retriever.from_documents(chunks)
        bm25_retriever.k = 10
        
        rag_pipeline["retriever"] = EnsembleRetriever(
            retrievers=[bm25_retriever, faiss_retriever],
            weights=[0.5, 0.5]
        )
        
        return f"File '{os.path.basename(file_path)}' berhasil diproses! Silakan ajukan pertanyaan.", gr.update(interactive=True)
    except Exception as e:
        return f"Error saat memproses file: {str(e)}", gr.update(interactive=False)


def get_rag_response(query, chat_history):
    if rag_pipeline["retriever"] is None:
        return "Dokumen belum diproses. Mohon unggah file terlebih dahulu."

    query_original = query
    query_lower = query_original.lower()
    final_answer = ""
    found_source = "Tidak ada sumber spesifik"

    priority_keywords = ["jumlah aset lancar"]
    use_smart_lane = any(keyword in query_lower for keyword in priority_keywords)

    if use_smart_lane:
        # Jalur Cerdas
        year_match = re.search(r'\b(202[3-4])\b', query_lower)
        target_year = year_match.group(1) if year_match else "2024"
        for chunk in rag_pipeline["all_chunks"]:
            lines = chunk.page_content.split('\n')
            for line in lines:
                if any(keyword in line.lower() for keyword in priority_keywords):
                    numbers = re.findall(r'(\d{1,3}(?:[.,]\d{3})*)', line)
                    if len(numbers) >= 2:
                        value_2024 = numbers[0]
                        value_2023 = numbers[1]
                        value = value_2024 if target_year == "2024" else value_2023
                        final_answer = f"Jumlah aset lancar untuk tahun {target_year} adalah **{value}**."
                        found_source = f"Sumber: Halaman {chunk.metadata.get('page', 'NA')}"
                        break
            if final_answer: break
    
    if not final_answer:
        # Jalur Normal
        retrieved_docs = rag_pipeline["retriever"].invoke(query_original)
        clean_context = "\n\n".join([doc.page_content for doc in retrieved_docs[:3]])
        found_source = ", ".join(list(set([f"Halaman {doc.metadata.get('page', 'NA')}" for doc in retrieved_docs[:3]])))
        
        chat_template = [{"role": "system", "content": "Anda adalah AI analis keuangan yang teliti. Jawab pertanyaan hanya berdasarkan teks yang diberikan."}, {"role": "user", "content": f"Dari TEKS di bawah, temukan jawaban untuk pertanyaan '{query_original}'.\n\nTEKS:\n{clean_context}\n\nJAWABAN:"}]
        final_prompt = tokenizer.apply_chat_template(chat_template, tokenize=False, add_generation_prompt=True)
        inputs = tokenizer(final_prompt, return_tensors="pt").to(llm.device)
        outputs = llm.generate(**inputs, max_new_tokens=250, do_sample=False, pad_token_id=tokenizer.eos_token_id)
        input_length = inputs.input_ids.shape[1]
        generated_tokens = outputs[0, input_length:]
        final_answer = tokenizer.decode(generated_tokens, skip_special_tokens=True)

    full_response = f"{final_answer}\n\n*{found_source}*"
    chat_history.append((query, full_response))
    return "", chat_history


# --- 3. MEMBUAT UI DENGAN GRADIO ---
with gr.Blocks() as demo:
    gr.Markdown("# 📊 Financial RAG Chatbot")
    
    with gr.Row():
        with gr.Column(scale=1):
            file_output = gr.Textbox(label="Status Dokumen", interactive=False)
            upload_button = gr.UploadButton("Klik untuk Upload PDF", file_types=[".pdf"])
            ask_button = gr.Button("Tanya", interactive=False)

        with gr.Column(scale=4):
            chatbot = gr.Chatbot(label="Chat")
            msg = gr.Textbox(label="Ketik Pertanyaan Anda di Sini...")
            
    # Hubungkan Aksi dengan Fungsi
    upload_button.upload(process_document, upload_button, [file_output, ask_button])
    msg.submit(get_rag_response, [msg, chatbot], [msg, chatbot])
    ask_button.click(get_rag_response, [msg, chatbot], [msg, chatbot])

# --- 4. JALANKAN APLIKASI ---
if __name__ == "__main__":
    demo.launch()