streamlit_app.py

import streamlit as st
import torch
import os
from transformers import AutoTokenizer, AutoModelForCausalLM
import traceback
from typing import Optional

# Configure the page
st.set_page_config(
    page_title="LLM Comparison: GPT-4 vs Gemini vs AOE",
    page_icon="⚔️",
    layout="wide"
)

def load_aoe_model():
    """Load the AoE model and tokenizer from outputs/student/ directory"""
    model_path = "outputs/student/"
    
    try:
        if not os.path.exists(model_path):
            st.error(f"Model directory '{model_path}' not found. Please ensure the model files are present.")
            return None, None
        
        # Check if required files exist
        required_files = ["config.json", "pytorch_model.bin", "tokenizer.json"]
        missing_files = [f for f in required_files if not os.path.exists(os.path.join(model_path, f))]
        
        if missing_files:
            st.warning(f"Some model files may be missing: {missing_files}. Attempting to load anyway...")
        
        # Load tokenizer and model
        tokenizer = AutoTokenizer.from_pretrained(model_path, trust_remote_code=True)
        model = AutoModelForCausalLM.from_pretrained(
            model_path, 
            torch_dtype=torch.float16 if torch.cuda.is_available() else torch.float32,
            device_map="auto" if torch.cuda.is_available() else None,
            trust_remote_code=True
        )
        
        return model, tokenizer
        
    except Exception as e:
        st.error(f"Error loading AoE model: {str(e)}")
        st.text(f"Traceback: {traceback.format_exc()}")
        return None, None

def generate_aoe_response(model, tokenizer, prompt, max_length=512):
    """Generate response from the AoE model"""
    try:
        # Tokenize input
        inputs = tokenizer.encode(prompt, return_tensors="pt")
        
        # Move to same device as model if CUDA is available
        if torch.cuda.is_available() and next(model.parameters()).is_cuda:
            inputs = inputs.cuda()
        
        # Generate response
        with torch.no_grad():
            outputs = model.generate(
                inputs,
                max_length=len(inputs[0]) + max_length,
                num_return_sequences=1,
                temperature=0.7,
                do_sample=True,
                pad_token_id=tokenizer.eos_token_id
            )
        
        # Decode response
        response = tokenizer.decode(outputs[0], skip_special_tokens=True)
        
        # Remove the input prompt from the response
        if response.startswith(prompt):
            response = response[len(prompt):].strip()
        
        return response
        
    except Exception as e:
        return f"Error generating AoE response: {str(e)}"

def query_gpt4_api(prompt: str) -> str:
    """Query GPT-4 API using environment variable for API key"""
    api_key = os.getenv('OPENAI_API_KEY')
    
    if not api_key:
        return "❌ GPT-4 API key not found in environment variables. Please set OPENAI_API_KEY environment variable to use GPT-4."
    
    try:
        # This is a placeholder implementation - would need actual OpenAI API integration
        return "🤖 GPT-4 response would appear here with proper API configuration."
    except Exception as e:
        return f"Error querying GPT-4: {str(e)}"

def query_gemini_api(prompt: str) -> str:
    """Query Gemini API using environment variable for API key"""
    api_key = os.getenv('GOOGLE_API_KEY')
    
    if not api_key:
        return "❌ Gemini API key not found in environment variables. Please set GOOGLE_API_KEY environment variable to use Gemini."
    
    try:
        # This is a placeholder implementation - would need actual Google Gemini API integration
        return "🤖 Gemini response would appear here with proper API configuration."
    except Exception as e:
        return f"Error querying Gemini: {str(e)}"

def main():
    st.title("⚔️ LLM Comparison: GPT-4 vs Gemini vs AOE")
    st.markdown("Compare responses from three different language models side by side.")
    
    # Initialize session state for model caching
    if 'aoe_model' not in st.session_state:
        st.session_state.aoe_model = None
        st.session_state.aoe_tokenizer = None
        st.session_state.aoe_loaded = False
    
    # Load AOE model on first run
    if not st.session_state.aoe_loaded:
        with st.spinner("Loading AOE model from outputs/student/..."):
            model, tokenizer = load_aoe_model()
            if model is not None and tokenizer is not None:
                st.session_state.aoe_model = model
                st.session_state.aoe_tokenizer = tokenizer
                st.session_state.aoe_loaded = True
                st.success("✅ AOE model loaded successfully!")
            else:
                st.error("❌ Failed to load AOE model. Check error messages above.")
    
    # Configuration section
    st.markdown("---")
    st.subheader("🔧 Configuration")
    
    col1, col2 = st.columns(2)
    
    with col1:
        max_length = st.slider(
            "Max Response Length",
            min_value=100,
            max_value=1000,
            value=512,
            step=50,
            help="Maximum length for generated responses"
        )
    
    with col2:
        # Display API key status
        openai_key_status = "✅ Found" if os.getenv('OPENAI_API_KEY') else "❌ Missing"
        google_key_status = "✅ Found" if os.getenv('GOOGLE_API_KEY') else "❌ Missing"
        
        st.info(f"**API Key Status:**\n\nOpenAI API Key: {openai_key_status}\n\nGoogle API Key: {google_key_status}")
    
    # Main comparison interface
    st.markdown("---")
    st.subheader("💬 Compare LLM Responses")
    
    # User input
    user_prompt = st.text_area(
        "Enter your prompt:",
        placeholder="Type your prompt here to compare responses from all three models...",
        height=120,
        help="Enter a prompt to see how different LLMs respond"
    )
    
    # Generate responses button
    if st.button("🚀 Generate All Responses", type="primary"):
        if not user_prompt.strip():
            st.warning("Please enter a prompt first.")
        else:
            # Create three columns for side-by-side comparison
            col1, col2, col3 = st.columns(3)
            
            with col1:
                st.markdown("### 🤖 GPT-4")
                with st.spinner("Generating GPT-4 response..."):
                    gpt4_response = query_gpt4_api(user_prompt)
                st.markdown("**Response:**")
                st.write(gpt4_response)
            
            with col2:
                st.markdown("### 🌟 Gemini")
                with st.spinner("Generating Gemini response..."):
                    gemini_response = query_gemini_api(user_prompt)
                st.markdown("**Response:**")
                st.write(gemini_response)
            
            with col3:
                st.markdown("### 🏰 AOE (Local)")
                if st.session_state.aoe_loaded:
                    with st.spinner("Generating AOE response..."):
                        aoe_response = generate_aoe_response(
                            st.session_state.aoe_model,
                            st.session_state.aoe_tokenizer,
                            user_prompt,
                            max_length
                        )
                    st.markdown("**Response:**")
                    st.write(aoe_response)
                else:
                    st.error("AOE model not loaded. Please reload the page.")
    
    # Model information sidebar
    with st.sidebar:
        st.header("ℹ️ Model Information")
        
        st.markdown("**🤖 GPT-4**")
        openai_status = "✅ Configured" if os.getenv('OPENAI_API_KEY') else "❌ Environment variable OPENAI_API_KEY not set"
        st.write(f"Status: {openai_status}")
        st.write("Provider: OpenAI")
        
        st.markdown("**🌟 Gemini**")
        google_status = "✅ Configured" if os.getenv('GOOGLE_API_KEY') else "❌ Environment variable GOOGLE_API_KEY not set"
        st.write(f"Status: {google_status}")
        st.write("Provider: Google")
        
        st.markdown("**🏰 AOE (Local)**")
        st.write(f"Status: {'✅ Loaded' if st.session_state.aoe_loaded else '❌ Not loaded'}")
        st.write("Path: outputs/student/")
        if st.session_state.aoe_loaded:
            try:
                device_info = f"Device: {next(st.session_state.aoe_model.parameters()).device}"
                st.write(device_info)
            except:
                pass
        
        if st.button("🔄 Reload AOE Model"):
            st.session_state.aoe_loaded = False
            st.experimental_rerun()
        
        st.markdown("---")
        st.markdown("**📋 Instructions:**")
        st.markdown("1. Set OPENAI_API_KEY and GOOGLE_API_KEY environment variables")
        st.markdown("2. Enter your prompt in the text area")
        st.markdown("3. Click 'Generate All Responses'")
        st.markdown("4. Compare responses side by side")
        
        st.markdown("---")
        st.markdown("**⚠️ Notes:**")
        st.markdown("- GPT-4 and Gemini require valid API keys in environment variables")
        st.markdown("- AOE model runs locally from outputs/student/")
        st.markdown("- Responses are generated independently")

if __name__ == "__main__":
    main()