services/chatbot_service.py

from transformers import pipeline, AutoTokenizer, AutoModelForCausalLM
import logging
import torch

logger = logging.getLogger(__name__)

# Global chatbot components
chatbot_pipeline = None
chat_history = {}

def load_chatbot_model():
    """Load the free DialoGPT model for chatbot"""
    global chatbot_pipeline
    try:
        logger.info("Loading DialoGPT chatbot model...")
        
        # Use DialoGPT medium for better responses
        chatbot_pipeline = pipeline(
            "text-generation",
            model="microsoft/DialoGPT-medium",
            tokenizer="microsoft/DialoGPT-medium",
            device="cpu"
        )
        logger.info("✓ DialoGPT chatbot model loaded successfully")
    except Exception as e:
        logger.error(f"✗ Failed to load DialoGPT model: {str(e)}")
        chatbot_pipeline = None

async def get_chatbot_response(user_text: str, user_id: str = "default") -> str:
    """
    Generate chatbot response using free DialoGPT model.
    
    Args:
        user_text: User input text
        user_id: Unique user ID for maintaining conversation history
    
    Returns:
        Chatbot response text
    """
    global chatbot_pipeline
    
    try:
        if chatbot_pipeline is None:
            load_chatbot_model()
            if chatbot_pipeline is None:
                return get_fallback_response(user_text)
        
        logger.info(f"Generating chatbot response for: '{user_text}'")
        
        # Get or initialize chat history for this user
        if user_id not in chat_history:
            chat_history[user_id] = []
        
        # Prepare conversation context
        conversation = chat_history[user_id] + [user_text]
        context = " ".join(conversation[-3:])  # Use last 3 exchanges as context
        
        # Generate response
        response = chatbot_pipeline(
            context,
            max_length=150,
            num_return_sequences=1,
            pad_token_id=chatbot_pipeline.tokenizer.eos_token_id,
            no_repeat_ngram_size=3,
            do_sample=True,
            top_k=50,
            top_p=0.95,
            temperature=0.7
        )
        
        bot_response = response[0]['generated_text'].strip()
        
        # Extract only the new response (remove the input context)
        if context in bot_response:
            bot_response = bot_response.replace(context, "").strip()
        
        # Clean up the response
        bot_response = clean_response(bot_response)
        
        # Update chat history
        chat_history[user_id].extend([user_text, bot_response])
        
        # Keep only recent history (last 4 exchanges)
        if len(chat_history[user_id]) > 8:
            chat_history[user_id] = chat_history[user_id][-8:]
        
        logger.info(f"✓ Response generated: '{bot_response}'")
        return bot_response
        
    except Exception as e:
        logger.error(f"✗ Chatbot response failed: {str(e)}")
        return get_fallback_response(user_text)


def clean_response(response: str) -> str:
    """Clean and format the chatbot response"""
    # Remove extra spaces
    response = ' '.join(response.split())
    
    # Ensure proper sentence ending
    if response and not response.endswith(('.', '!', '?')):
        response += '.'
    
    # Limit response length
    if len(response) > 200:
        response = response[:197] + '...'
    
    return response


def get_fallback_response(user_text: str) -> str:
    """Provide fallback responses when model fails"""
    fallback_responses = [
        f"I understand you said: '{user_text}'. Could you tell me more about that?",
        f"That's interesting! You mentioned: '{user_text}'. What would you like to know?",
        f"Thanks for sharing! Regarding '{user_text}', how can I help you?",
        f"I heard you say: '{user_text}'. Could you elaborate on that?"
    ]
    
    import random
    return random.choice(fallback_responses)