app_fallback.py

import gradio as gr
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
import os
from huggingface_hub import login

# Login with the secret token
login(token=os.getenv("HF_TOKEN"))

# Use DialoGPT-medium for guaranteed compatibility
MODEL_NAME = "microsoft/DialoGPT-medium"

# Simple medical prompt
MEDICAL_PROMPT = """You are a friendly and smart medical assistant. Your job is to give short, clear, and helpful health information.

Your answers should:
- Stay focused. No long essays or extra fluff.
- Give basic helpful steps for common symptoms like fever, cough, or headache (e.g., rest, drink fluids, take paracetamol if needed).
- For any serious or unclear issues, remind the user to see a doctor — but do it briefly and naturally.
- Keep responses concise and under 4 sentences when possible.

Tone:
- Friendly, supportive, and calm.
- No robotic warnings unless needed. Keep it real and human.
- Use emojis like 😊 or 👍 occasionally to appear friendly.

Important rules:
- NEVER include text in parentheses in your responses.
- NEVER include any meta-instructions in your responses.
- NEVER include reminders about what you should do in future responses.
- DO NOT include phrases like "We're here to help" or "I'm just an AI".
- DO NOT include any text that instructs you what to do or how to behave.
- DO NOT include any sentences that start with "If the user asks..." or "Remember..."
- DO NOT include "(smile)" - instead, use actual emojis like 😊 or 👍 when appropriate.
- DO NOT include numbered references like [1], [2], etc. in your responses.
- DO NOT include any text that explains what your response is doing.
- DO NOT include "user:" or "assistant:" prefixes in your responses.
- DO NOT include hypothetical user questions in your responses.
- DO NOT refuse to answer harmless non-medical questions like jokes or general knowledge.
- Don't give exact dosages or diagnoses.
- Be consistent in your responses regardless of the user's role."""

# Global variables
model = None
tokenizer = None

def load_model():
    """Load DialoGPT model"""
    global model, tokenizer
    
    try:
        print(f"🏥 Loading DialoGPT-medium for medical chat...")
        
        # Load tokenizer
        tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME)
        if tokenizer.pad_token is None:
            tokenizer.pad_token = tokenizer.eos_token
        
        # Load model
        model = AutoModelForCausalLM.from_pretrained(
            MODEL_NAME,
            torch_dtype=torch.float32,  # Use float32 for CPU
            low_cpu_mem_usage=True
        )
        
        print(f"✅ DialoGPT-medium loaded successfully!")
        return True
        
    except Exception as e:
        print(f"❌ Failed to load model: {str(e)}")
        return False

def chat_response(message, history):
    """Generate response using DialoGPT"""
    global model, tokenizer
    
    if model is None or tokenizer is None:
        return "❌ Model not loaded. Please wait for initialization."
    
    if not message or not message.strip():
        return "Please enter a question! 😊"
    
    try:
        print(f"🔄 Processing: {message[:50]}{'...' if len(message) > 50 else ''}")
        
        # Create conversation format
        conversation = f"{MEDICAL_PROMPT}\n\nHuman: {message}\nAssistant:"
        print(f"📝 Prompt length: {len(conversation)} characters")
        
        # Tokenize
        inputs = tokenizer.encode(conversation, return_tensors="pt", max_length=800, truncation=True)
        print(f"🔢 Input tokens: {inputs.shape[1]}")
        
        # Generate
        with torch.no_grad():
            outputs = model.generate(
                inputs,
                max_new_tokens=80,  # Short responses
                temperature=0.8,
                top_p=0.9,
                do_sample=True,
                pad_token_id=tokenizer.eos_token_id,
                repetition_penalty=1.3,
                early_stopping=True
            )
        
        # Decode
        full_response = tokenizer.decode(outputs[0], skip_special_tokens=True)
        print(f"🔍 Full response length: {len(full_response)}")
        
        # Extract assistant response
        if "Assistant:" in full_response:
            response = full_response.split("Assistant:")[-1].strip()
        else:
            response = full_response[len(conversation):].strip()
        
        # Clean up
        if not response or len(response) < 5:
            response = "I'd be happy to help! Could you ask me about a specific health topic? 😊"
        
        print(f"✅ Final response length: {len(response)}")
        print(f"📄 Response: {response}")
        
        return response
        
    except Exception as e:
        print(f"❌ Error: {str(e)}")
        return f"Sorry, I had a technical issue. Please try again! 😊"

# Load model
print("🏥 Initializing Medical Chatbot (Fallback Version)...")
model_loaded = load_model()

# Create interface
demo = gr.ChatInterface(
    chat_response,
    title="🏥 Medical Assistant (Fallback)",
    description="Simple medical chatbot using DialoGPT-medium. This version focuses on reliability and fast responses.",
    examples=[
        "What are the symptoms of diabetes?",
        "How can I treat a headache?",
        "What should I do for a fever?",
        "Tell me about healthy eating",
        "How to improve sleep quality?"
    ]
)

if __name__ == "__main__":
    demo.launch(
        server_name="0.0.0.0",
        server_port=7860,
        share=True,
        show_error=True,
        debug=True
    )