app.py

import gradio as gr
import torch
from transformers import AutoTokenizer, AutoModelForMaskedLM
import wikipedia
import speech_recognition as sr

# ✅ নতুন Model
MODEL_NAME = "csebuetnlp/banglabert"
tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME)
model = AutoModelForMaskedLM.from_pretrained(MODEL_NAME)

memory = {}

# ✅ Wikipedia search
def search_wiki(query):
    try:
        result = wikipedia.summary(query, sentences=2)
        return result
    except:
        return None

# ✅ প্রশ্ন-উত্তর সিস্টেম
def generate_response(user_input):
    inputs = tokenizer(user_input, return_tensors="pt")
    outputs = model(**inputs)
    logits = outputs.logits
    predicted_index = torch.argmax(logits[0], dim=-1)
    decoded = tokenizer.decode(predicted_index)
    return decoded.strip()

# ✅ চ্যাটবট লজিক
def sujata_chat(user_input):
    if user_input.startswith("শেখাও:"):
        try:
            parts = user_input.replace("শেখাও:", "").split("=")
            q, a = parts[0].strip(), parts[1].strip()
            memory[q] = a
            return f"✅ শেখা সম্পন্ন:\n❓ {q}\n✅ {a}"
        except:
            return "ভুল ফরম্যাট। লিখুন: শেখাও: প্রশ্ন = উত্তর"

    elif user_input in memory:
        return memory[user_input]

    elif "কেন" in user_input or "হয়" in user_input:
        return "এটি একটি যুক্তির প্রশ্ন। আমি ব্যাখ্যা করতে চেষ্টা করছি..."

    result = search_wiki(user_input)
    if result:
        return result

    return generate_response(user_input)

# ✅ ভয়েস টু টেক্সট
def voice_to_text(audio):
    recognizer = sr.Recognizer()
    with sr.AudioFile(audio) as source:
        audio_data = recognizer.record(source)
    try:
        return recognizer.recognize_google(audio_data, language="bn-BD")
    except:
        return "আমি বুঝতে পারিনি, আবার বলুন।"

# ✅ UI
with gr.Blocks(css="body {background-color: #121212; color: white;}") as demo:
    gr.Markdown("<h2 style='text-align: center; color:white;'>🌙 সুজাতা AI (বাংলা চ্যাটবট)</h2>")

    chatbot = gr.Chatbot(label="সুজাতা", bubble_full_width=False)
    msg = gr.Textbox(label="✍️ প্রশ্ন করুন", placeholder="এখানে বাংলায় লিখুন...", lines=2)

    with gr.Row():
        send = gr.Button("📤 পাঠাও")
        clear = gr.Button("🗑️ মুছুন")

    with gr.Row():
        mic_input = gr.Audio(source="microphone", type="filepath", label="🎙️ কণ্ঠস্বর ইনপুট")
        mic_btn = gr.Button("🎧 পাঠাও")

    def respond(user_message, chat_history):
        output = sujata_chat(user_message)
        chat_history.append((user_message, output))
        return chat_history, ""

    def respond_from_voice(audio):
        user_text = voice_to_text(audio)
        output = sujata_chat(user_text)
        return [(user_text, output)]

    send.click(respond, [msg, chatbot], [chatbot, msg])
    mic_btn.click(respond_from_voice, mic_input, chatbot)
    clear.click(lambda: [], None, chatbot)

demo.launch()