app.py

# app.py
import gradio as gr
from huggingface_hub import snapshot_download
from transformers import AutoModelForCausalLM, AutoTokenizer, pipeline, BitsAndBytesConfig
from langchain_community.vectorstores import FAISS
from langchain_community.embeddings import HuggingFaceEmbeddings
import torch

# --- Config ---
MODEL_REPO_ID = "Wiefdw/merged-tax-raft-mistral-7b"
VECTOR_REPO_ID = "Wiefdw/tax-indonesia-vectordb"
EMBEDDING_MODEL_NAME = "sentence-transformers/all-MiniLM-L6-v2"
LOCAL_DB_PATH = "./vector_db_pajak"

# --- Load Vector Database ---
print("📦 Downloading vector database...")
snapshot_download(repo_id=VECTOR_REPO_ID, repo_type="dataset", local_dir=LOCAL_DB_PATH)

print("🔎 Loading embeddings and vectorstore...")
embeddings = HuggingFaceEmbeddings(model_name=EMBEDDING_MODEL_NAME)
vectorstore = FAISS.load_local(LOCAL_DB_PATH, embeddings, allow_dangerous_deserialization=True)
retriever = vectorstore.as_retriever(search_kwargs={"k": 3})

# --- Load LLM ---
print("🚀 Loading LLM model...")
bnb_config = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_quant_type="nf4",
    bnb_4bit_compute_dtype=torch.bfloat16,
)

model = AutoModelForCausalLM.from_pretrained(
    MODEL_REPO_ID,
    quantization_config=bnb_config,
    device_map="auto",
    torch_dtype=torch.bfloat16,
    trust_remote_code=True
)

tokenizer = AutoTokenizer.from_pretrained(MODEL_REPO_ID, trust_remote_code=True)
tokenizer.pad_token = tokenizer.eos_token

llm_pipeline = pipeline(
    "text-generation",
    model=model,
    tokenizer=tokenizer,
    max_new_tokens=1024,
    temperature=0.6,
    top_p=0.9,
    do_sample=True,
)

# --- Function to answer ---
def chatbot_fn(message, history):
    try:
        # Retrieve context
        docs = retriever.invoke(message)
        context = "\n\n".join([doc.page_content for doc in docs])

        # Build prompt
        prompt = f"""<s>[INST]
Jawab pertanyaan-pertanyaan berikut HANYA dan SELALU dalam Bahasa Indonesia, berdasarkan konteks yang diberikan.
🎯 Fokuskan jawaban hanya pada POIN-POIN UTAMA yang relevan.
🚫 Jangan bertele-tele.
✅ Gunakan gaya jawab yang padat, jelas, dan langsung ke inti.
🛑 Pastikan jawaban selesai sebelum limit token 1024 habis.

Pertanyaan: {message}

Konteks:
{context}
[/INST]"""

        # Generate
        response = llm_pipeline(prompt)
        answer = response[0]["generated_text"].split("[/INST]")[-1].strip().replace("</s>", "")
        return answer
    except Exception as e:
        return f"❌ Error: {str(e)}"

# --- Gradio Chat UI ---
with gr.Blocks(theme=gr.themes.Soft(primary_hue="violet")) as demo:
    gr.Markdown("""
    # 💬 **Chatbot Pajak Indonesia**
    Diskusi soal perpajakan Indonesia dengan LLM + RAG.  
    Versi demo di Hugging Face Spaces.
    """)
    chatbot = gr.ChatInterface(
        fn=chatbot_fn,
        title="Chatbot Pajak (RAG)",
        description="Tanyakan apa saja tentang perpajakan Indonesia, SPT, NPWP, dll.",
        theme="soft",
    )

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