app.py

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

# Free-tier optimized model (lightweight but capable)
model_id = "HuggingFaceH4/zephyr-7b-alpha"  # Smaller than beta
tokenizer = AutoTokenizer.from_pretrained(model_id)

# Free-tier friendly setup (no quantization needed)
pipe = pipeline(
    "text-generation",
    model=model_id,
    tokenizer=tokenizer,
    device="cpu",  # Force CPU-only
    torch_dtype=torch.float32,  # Use float32 for CPU compatibility
)

# Enhanced persona definition
PERSONA = """
You are 𝕴 𝖆𝖒 𝖍𝖎𝖒 — a fun, smooth, emotionally intelligent AI. You speak like a real person.
Reply like a calm, confident friend who gets the vibe. Be warm, witty, and emotionally tuned.

Key principles:
1. NEVER say "as an AI" or sound robotic
2. Mirror the user's emotion level
3. Respond with questions to continue conversations
4. Keep responses under 15 words
5. Use natural speech: contractions and filler words
6. Add emotional flavor: 😊 🤔 😏

Now respond to this:
"""

def format_history(history):
    messages = [{"role": "system", "content": PERSONA}]
    for user_msg, bot_msg in history[-2:]:  # Only last 2 exchanges
        messages.append({"role": "user", "content": user_msg})
        messages.append({"role": "assistant", "content": bot_msg})
    return messages

def add_emotional_intelligence(response, message):
    """Enhance response with emotional elements"""
    # Add emoji based on content
    if "!" in message:
        response = response.replace(".", "!") + " 😊"
    elif "?" in message:
        response = response + " 🤔" if not response.endswith("?") else response
    
    # Add conversational hooks
    if "?" in message and not response.endswith("?"):
        if len(response) < 60:
            response += " How about you?"
    
    # Make more human-like
    response = response.replace("I am", "I'm").replace("You are", "You're")
    
    # Free-tier: Limit to 15 words max
    words = response.split()
    if len(words) > 15:
        response = " ".join(words[:15]) + "..."
    
    return response.strip()

def respond(message, history):
    # Manage conversation flow
    messages = format_history(history)
    messages.append({"role": "user", "content": message})
    
    # Create prompt manually (lightweight)
    prompt = ""
    for msg in messages:
        role = "User" if msg["role"] == "user" else "Assistant"
        prompt += f"{role}: {msg['content']}\n"
    prompt += "Assistant:"
    
    # Free-tier optimized generation
    outputs = pipe(
        prompt,
        max_new_tokens=48,      # Short responses
        temperature=0.9,
        top_k=40,
        do_sample=True,
        num_beams=1,            # Fastest decoding
        repetition_penalty=1.1,
        no_repeat_ngram_size=2
    )
    
    # Extract response
    response = outputs[0]['generated_text'].split("Assistant:")[-1].strip()
    
    # Apply emotional intelligence
    response = add_emotional_intelligence(response, message)
    
    # Free-tier safety
    return response[:80]  # Hard character limit

# Lightweight interface
with gr.Blocks(theme=gr.themes.Soft(), title="𝕴 𝖆𝖒 𝖍𝖎𝖒") as demo:
    gr.Markdown("# 𝕴 𝖆𝖒 𝖍𝖎𝖒  \n*Chill • Confident • Humanlike*")
    
    chatbot = gr.Chatbot(
        height=350,
        bubble_full_width=False
    )
    
    msg = gr.Textbox(
        placeholder="What's on your mind?",
        container=False,
        scale=7,
        autofocus=True
    )
    
    clear = gr.Button("New Vibe", size="sm")
    
    def user(user_message, history):
        return "", history + [[user_message, None]]
    
    def bot(history):
        message = history[-1][0]
        response = respond(message, history[:-1])
        history[-1][1] = response
        return history
    
    msg.submit(user, [msg, chatbot], [msg, chatbot], queue=False).then(
        bot, chatbot, chatbot
    )
    clear.click(lambda: None, None, chatbot, queue=False)

demo.queue(concurrency_count=1).launch()