app.py

import torch
from transformers import AutoModelForSeq2SeqLM, AutoTokenizer
import gradio as gr
import re

# ---------------------------------------------------------
# 1️⃣ Load Model
# ---------------------------------------------------------
torch.set_num_threads(1)
print("🤖 Starting Model Loading...")

try:
    # Use small version for faster inference and better generalization
    MODEL_NAME = "google/mt5-small"
    TOKENIZER = AutoTokenizer.from_pretrained(MODEL_NAME)
    CHAT_MODEL = AutoModelForSeq2SeqLM.from_pretrained(MODEL_NAME)
    print("✅ Model loaded successfully.")
except Exception as e:
    print(f"❌ Model loading failed: {e}")
    TOKENIZER, CHAT_MODEL = None, None


# ---------------------------------------------------------
# 2️⃣ Detect Script (Sindhi or Roman Sindhi)
# ---------------------------------------------------------
def detect_script(text):
    """Detect whether text is Sindhi (Arabic script) or Roman Sindhi (Latin script)."""
    if re.search(r'[\u0600-\u06FF]', text):
        return "sindhi"
    else:
        return "roman"


# ---------------------------------------------------------
# 3️⃣ Chat Function
# ---------------------------------------------------------
def generate_reply(user_text):
    """Generate Sindhi or Roman Sindhi reply based on input language."""
    if not user_text.strip():
        return "مهرباني ڪري ڪجھ لکو."  # Please type something.

    script_type = detect_script(user_text)

    # Improved instruction-style prompts
    if script_type == "sindhi":
        prompt = f"توھان جو ڪم ھي آھي ته ھيٺين جملي جو جواب سنڌي ۾ ڏيو:\nسوال: {user_text}\nجواب:"
    else:
        prompt = f"Tuhanjo kaam aahe ta neeche likhe sawal jo jawab Roman Sindhi mein likho:\nSawaal: {user_text}\nJawab:"

    inputs = TOKENIZER(prompt, return_tensors="pt", truncation=True)
    outputs = CHAT_MODEL.generate(
        **inputs,
        max_new_tokens=100,
        do_sample=True,
        top_p=0.95,
        top_k=40,
        temperature=0.8
    )

    reply = TOKENIZER.decode(outputs[0], skip_special_tokens=True)

    # Remove unwanted tokens (like <extra_id_0>, etc.)
    reply = re.sub(r"<.*?>", "", reply).strip()
    return reply


# ---------------------------------------------------------
# 4️⃣ Gradio Interface
# ---------------------------------------------------------
with gr.Blocks(title="Sindhi Text Chatbot") as app:
    gr.Markdown(
        """
        # 💬 سنڌي چيٽ بوٽ (Text Only)
        ✍️ رومن سنڌي يا سنڌي ۾ لکو — ۽ بوٽ اوهان کي ساڳئي ٻولي ۾ جواب ڏيندو.  
        **Example:**  
        - Sindhi: توھان جو نالو ڇا آهي؟  
        - Roman Sindhi: Tuhanjo naalo chaa aahe?
        """
    )

    with gr.Row():
        with gr.Column(scale=1):
            user_input = gr.Textbox(
                label="✍️ پنهنجو سوال لکو | Type your message (Sindhi or Roman Sindhi)",
                lines=3,
                placeholder="مثال طور: توھان جو نالو ڇا آهي؟ يا Tuhanjo naalo chaa aahe?"
            )
            send_btn = gr.Button("Send", variant="primary")

        with gr.Column(scale=2):
            reply_output = gr.Textbox(
                label="💬 AI جو جواب | AI’s Reply",
                lines=5
            )
            clear_btn = gr.Button("Clear")

    # Main button action
    send_btn.click(
        fn=generate_reply,
        inputs=[user_input],
        outputs=[reply_output]
    )

    # Clear all fields
    def clear_all():
        return "", ""

    clear_btn.click(
        fn=clear_all,
        inputs=[],
        outputs=[user_input, reply_output]
    )


# ---------------------------------------------------------
# 5️⃣ Launch App
# ---------------------------------------------------------
if __name__ == "__main__":
    app.launch()