app.py

import gradio as gr
import random
import pandas as pd
import os
import threading
import time
import numpy as np
from utils.data_loader import get_random_example
from utils.models import generate_summaries, model_names
from utils.ui_helpers import toggle_context_display, update_feedback, get_context_html, toggle_reference_answer
from utils.leaderboard import load_leaderboard_data, submit_vote_with_elo, generate_leaderboard_html
from utils.vote_logger import save_vote_details
from utils.shared import generation_interrupt

feedback_options = {
    "left": [
        "Model A: Answers the question completely",
        "Model A: Information is accurate and correct", 
        "Model A: Stays on topic and relevant",
        "Model A: Clear and well-written response",
        "Model A: Appropriately says 'I don't know' without enough info",
        "Model A: Asks helpful follow-up questions when unclear"
    ],
    "right": [
        "Model B: Answers the question completely",
        "Model B: Information is accurate and correct",
        "Model B: Stays on topic and relevant", 
        "Model B: Clear and well-written response",
        "Model B: Appropriately says 'I don't know' without enough info",
        "Model B: Asks helpful follow-up questions when unclear"
    ],
    "tie": [
        "Model A: Answers the question completely",
        "Model A: Information is accurate and correct",
        "Model A: Stays on topic and relevant",
        "Model A: Clear and well-written response",
        "Model A: Appropriately says 'I don't know' without enough info",
        "Model A: Asks helpful follow-up questions when unclear",
        "Model B: Answers the question completely",
        "Model B: Information is accurate and correct",
        "Model B: Stays on topic and relevant",
        "Model B: Clear and well-written response",
        "Model B: Appropriately says 'I don't know' without enough info",
        "Model B: Asks helpful follow-up questions when unclear"
    ],
    "neither": [
        "Model A: Incomplete or missing key information",
        "Model A: Contains incorrect or made-up information",
        "Model A: Goes off-topic or irrelevant",
        "Model A: Poorly written or confusing",
        "Model A: Should have admitted uncertainty without enough info",
        "Model A: Should have asked clarifying questions but didn't",
        "Model B: Incomplete or missing key information", 
        "Model B: Contains incorrect or made-up information",
        "Model B: Goes off-topic or irrelevant",
        "Model B: Poorly written or confusing",
        "Model B: Should have admitted uncertainty without enough info",
        "Model B: Should have asked clarifying questions but didn't"
    ]
}

def weighted_sample_without_replacement(population, weights, k=2):
    """
    Performs a weighted random sampling without replacement.
    
    Args:
        population: The list of items to sample from
        weights: The weight for each item
        k: Number of items to sample
        
    Returns:
        A list of k sampled items
    """
    if len(population) <= k:
        return population
    
    # Convert weights to numpy array for efficient operations
    weights = np.array(weights)
    
    # Create a copy of the population and weights
    remaining_population = population.copy()
    remaining_weights = weights.copy()
    
    selected = []
    
    for _ in range(k):
        # Normalize weights so they sum to 1
        normalized_weights = remaining_weights / remaining_weights.sum()
        
        # Randomly select one item based on weights
        selected_idx = np.random.choice(len(remaining_population), p=normalized_weights)
        
        # Add the selected item to our result
        selected.append(remaining_population[selected_idx])
        
        # Remove the selected item from the pool
        remaining_population.pop(selected_idx)
        remaining_weights = np.delete(remaining_weights, selected_idx)
        
    return selected

def load_context(set_interrupt=False):
    if set_interrupt:
        generation_interrupt.set()
        time.sleep(0.2)
        
    generation_interrupt.clear()
    example = get_random_example()
    
    context_desc = example.get('processed_context_desc', '')
    if context_desc:
        context_desc = f"<div class='context-topic'><span class='topic-label'>The question and context are about:</span> {context_desc}</div>"
    
    show_full = False
    context_html = get_context_html(example, show_full=show_full)
    
    return [
        example,
        gr.update(value=example['question'], elem_classes="query-text"),  # Regular query styles
        gr.update(value=context_desc, visible=bool(context_desc)),
        gr.update(value=context_html),
        gr.update(value="Show Full Context", elem_classes=["context-toggle-button"], visible=True),  # Ensure toggle is visible
        show_full
    ]

def toggle_faq(expanded):
    """Toggle FAQ visibility with proper arrow icons"""
    new_state = not expanded
    button_text = "▼ Why can't I upload a file or ask my own question?" if new_state else "▶ Why can't I upload a file or ask my own question?"
    return new_state, gr.update(visible=new_state), gr.update(value=button_text)

# Explicit function to hide the FAQ section completely
def hide_faq_section():
    """Completely hide the FAQ section and its content"""
    return gr.update(visible=False), gr.update(visible=False)

def load_leaderboard():
    results = load_leaderboard_data()
    leaderboard_html = generate_leaderboard_html(results)
    return leaderboard_html

def generate_model_summaries(example):
    result = {
        "model_a": "",
        "model_b": "",
        "summary_a": "",
        "summary_b": "",
        "completed": False
    }
    
    if generation_interrupt.is_set():
        return result

    try:
        # Get current leaderboard data to determine model usage counts
        leaderboard_data = load_leaderboard_data()
        
        # Calculate weights using inverse weighting
        # Weight = K / (games_played + C)
        K = 100  # Scaling factor
        C = 5    # Smoothing constant
        
        weights = []
        model_list = []
        
        for model in model_names:
            # Get games played for the model, default to 0 if not found
            games_played = leaderboard_data["games_played"].get(model, 0)
            
            # Calculate weight using inverse formula
            weight = K / (games_played + C)
            
            weights.append(weight)
            model_list.append(model)
        
        # Select two models using weighted sampling without replacement
        selected_models = weighted_sample_without_replacement(model_list, weights, k=2)
        m_a_name, m_b_name = selected_models
        
        result["model_a"] = m_a_name
        result["model_b"] = m_b_name
        
        s_a, s_b = generate_summaries(example, m_a_name, m_b_name)
        
        if not generation_interrupt.is_set():
            result["summary_a"] = s_a
            result["summary_b"] = s_b
            result["completed"] = bool(s_a and s_b)
    except Exception as e:
        print(f"Error in generation: {e}")
        
    return result

def process_generation_result(result):
    if not result["completed"] or not result["summary_a"] or not result["summary_b"]:
        return [
            result.get("model_a", ""), 
            result.get("model_b", ""), 
            result.get("summary_a", ""), 
            result.get("summary_b", ""),
            None, [], False, load_leaderboard_data(),
            gr.update(value=result.get("summary_a", "Generation was interrupted or failed.")),
            gr.update(value=result.get("summary_b", "Generation was interrupted or failed.")),
            gr.update(interactive=False, elem_classes=["vote-button"]),
            gr.update(interactive=False, elem_classes=["vote-button"]),
            gr.update(interactive=False, elem_classes=["vote-button"]),
            gr.update(interactive=False, elem_classes=["vote-button", "vote-button-neither"]),
            gr.update(choices=[], value=[], interactive=False, visible=False),
            gr.update(visible=False),
            gr.update(interactive=False, visible=True),
            gr.update(visible=False),
            gr.update(interactive=True),
            gr.update(elem_classes=[])
        ]
    
    buttons_interactive = bool(result["summary_a"] and result["summary_b"])
    
    agg_results = load_leaderboard_data()
    return [
        result["model_a"], result["model_b"], 
        result["summary_a"], result["summary_b"],
        None, [], False, agg_results,
        gr.update(value=result["summary_a"]),
        gr.update(value=result["summary_b"]),
        gr.update(interactive=buttons_interactive, elem_classes=["vote-button"]),
        gr.update(interactive=buttons_interactive, elem_classes=["vote-button"]),
        gr.update(interactive=buttons_interactive, elem_classes=["vote-button"]),
        gr.update(interactive=buttons_interactive, elem_classes=["vote-button", "vote-button-neither"]),
        gr.update(choices=[], value=[], interactive=False, visible=False),
        gr.update(visible=False),
        gr.update(interactive=False, visible=True),
        gr.update(visible=False),
        gr.update(interactive=True),
        gr.update(elem_classes=[])
    ]

def process_example(example):
    result = generate_model_summaries(example)
    return process_generation_result(result)

def select_vote_improved(winner_choice):
    feedback_choices = feedback_options.get(winner_choice, [])

    btn_a_classes = ["vote-button"]
    btn_b_classes = ["vote-button"]
    btn_tie_classes = ["vote-button"]
    btn_neither_classes = ["vote-button", "vote-button-neither"]
    
    if winner_choice == 'left':
        btn_a_classes.append("selected")
    elif winner_choice == 'right':
        btn_b_classes.append("selected")
    elif winner_choice == 'tie':
        btn_tie_classes.append("selected")
    elif winner_choice == 'neither':
        btn_neither_classes.append("selected")

    return [
        winner_choice,
        gr.update(choices=feedback_choices, value=[], interactive=True, visible=True),
        gr.update(visible=True),
        gr.update(interactive=True),
        gr.update(elem_classes=btn_a_classes),
        gr.update(elem_classes=btn_b_classes),
        gr.update(elem_classes=btn_tie_classes),
        gr.update(elem_classes=btn_neither_classes)
    ]

def handle_vote_submission(example, m_a, m_b, winner, feedback, summary_a, summary_b, current_results):
    if winner is None:
        print("Warning: Submit called without a winner selected.")
        return {}

    save_vote_details(example, m_a, m_b, winner, feedback, summary_a, summary_b)
    return submit_vote_with_elo(m_a, m_b, winner, feedback, current_results)

def show_loading_state():
    """Show loading state while fetching new content and reset UI elements"""
    return [
        gr.update(value="Loading new question and summaries...", interactive=False),
        gr.update(value="Loading new question and summaries...", interactive=False),
        gr.update(interactive=False, elem_classes=["vote-button"]),  # Reset styling
        gr.update(interactive=False, elem_classes=["vote-button"]),
        gr.update(interactive=False, elem_classes=["vote-button"]),
        gr.update(interactive=False, elem_classes=["vote-button", "vote-button-neither"]),
        gr.update(visible=False),      # feedback_section
        gr.update(interactive=False),  # submit_button
        gr.update(visible=False),      # results_reveal_area
        gr.update(interactive=True),   # random_question_btn - KEEP ACTIVE during inference
        None  # Reset selected_winner
    ]

def handle_new_example_click():
    return load_context(set_interrupt=True)[0]

def update_ui_for_new_context(example):
    context_desc = example.get('processed_context_desc', '')
    if context_desc:
        context_desc = f"<div class='context-topic'><span class='topic-label'>The question and context are about:</span> {context_desc}</div>"
    
    return [
        gr.update(value=example['question'], elem_classes="query-text"),  # Regular query styles
        gr.update(value=context_desc, visible=bool(context_desc)),
        gr.update(value=get_context_html(example, False)),
        gr.update(value="Show Full Context", elem_classes=["context-toggle-button"], visible=True),  # Ensure toggle is visible
        False
    ]

def reset_reference_section():
    """Reset reference answer section to hidden state when loading new question"""
    return [
        False,  # Reset show_reference_answer state to False
        gr.update(visible=False),  # Hide reference content (like FAQ)
        gr.update(value="▶ Show Reference Answer"),  # Reset button text (like FAQ)
        gr.update(value="")  # Clear reference content
    ]

def cleanup_on_disconnect():
    print(f"Browser disconnected. Cleaning up resources...")
    generation_interrupt.set()

# Helper functions for showing/hiding UI elements
def initialize_empty_app():
    return [
        gr.update(visible=False),  # context_section
        gr.update(visible=False),  # model_section
        gr.update(visible=False),  # voting_section
        gr.update(visible=False)   # submit_button
    ]

def show_all_after_loading():
    return [
        gr.update(visible=True),  # context_section
        gr.update(visible=True),  # model_section
        gr.update(visible=True),  # voting_section
        gr.update(visible=True),  # submit_button
        gr.update(value="🔄 Try a New Question", elem_classes=["query-button"])  # update button text
    ]

with gr.Blocks(theme=gr.themes.Default(
    primary_hue=gr.themes.colors.orange,
    secondary_hue=gr.themes.colors.slate
)) as demo:
    css_path = os.path.join(os.getcwd(), 'static', 'styles.css')
    
    with open(css_path, 'r') as f:
        css_content = f.read()
    
    gr.HTML(f"<style>{css_content}</style>")
    
    unload_js = """
    <script>
    window.addEventListener('beforeunload', function(e) {
        navigator.sendBeacon('/cleanup?session_id=' + window.gradioClientState.session_hash);
    });
    </script>
    """
    gr.HTML(unload_js)

    # State variables
    current_example = gr.State({})
    model_a_name = gr.State("")
    model_b_name = gr.State("")
    summary_a_text = gr.State("")
    summary_b_text = gr.State("")
    selected_winner = gr.State(None)
    feedback_list = gr.State([])
    show_results_state = gr.State(False)
    results_agg = gr.State(load_leaderboard_data())
    show_full_context = gr.State(False)
    show_reference_answer = gr.State(False)  # NEW: State for reference answer toggle
    faq_expanded = gr.State(False)  # State for FAQ toggle

    with gr.Tabs() as tabs:
        with gr.TabItem("Arena", id="arena-tab"):
            gr.Markdown("# SLM RAG Arena -  Compare and Find The Best Sub-5B Models for RAG")
            gr.Markdown("""
🏟️ This arena evaluates how well small language models (under 5B) answer questions based on document contexts.

📝 Instructions：
-  **Click the "Get a Question" button** to load a random question with context
-  **Review the query and context** to understand the information provided to the models
-  **Compare answers** generated by two different models on answer quality or appropriate refusal
-  **Cast your vote** for the better response, or select 'Tie' if equally good or 'Neither' if both are inadequate
""")
            gr.Markdown("---")
            with gr.Column(elem_id="main-interface-area") as main_interface_area:
                with gr.Row(elem_id="query-title-row"):
                    gr.Markdown("### 💬 Query - Question About Document Content", elem_classes="section-heading")

                with gr.Row(elem_id="query-container"):
                    with gr.Row(elem_classes="query-box-row"):
                        query_display = gr.Markdown(value="Click \"Get a Question\" to start", elem_classes=["query-text", "empty-query"], elem_id="query-section")
                    random_question_btn = gr.Button("💡 Get a Question", elem_classes=["query-button", "initial-button"])
                
                # Add the FAQ toggle and content here
                with gr.Row(visible=True, elem_id="faq-container") as faq_container:
                    faq_toggle_btn = gr.Button("▶ Why can't I upload a file or ask my own question?", elem_classes=["faq-toggle-button"])
                
                # FAQ Content - initially hidden
                with gr.Row(visible=False, elem_id="faq-content") as faq_content:
                    gr.Markdown("""
                    This arena tests how well different AI models summarize information using standardized questions and contexts. All models see the exact same inputs for fair comparison.
                    
                    We don't allow file uploads here as that would change what we're measuring. Instead, check our leaderboard to find top-performing models for your needs. We'll soon launch a separate playground where you can test models with your own files.
                    """, elem_classes="faq-text")
                
                context_description = gr.Markdown("", elem_classes="context-description")
                
                # Create a section container for all context-related elements - INITIALLY HIDDEN
                with gr.Column(visible=False, elem_id="context-section") as context_section:
                    context_divider = gr.HTML("<hr>", elem_id="context-divider")
                    
                    with gr.Row(elem_id="context-header-row"):
                        gr.Markdown("### 📋 Context - Retrieved Content from the Document", elem_classes="context-title")
                        context_toggle_btn = gr.Button("Show Full Context", elem_classes=["context-toggle-button"])
                    
                    context_display = gr.HTML(value="", label="Context Chunks")

                # Model comparison section - initially hidden
                with gr.Column(visible=False, elem_id="model-section") as model_section:
                    gr.Markdown("---")
                    
                    # NEW: Model comparison header (simple)
                    gr.Markdown("### 🔍 Compare Models - Are these Grounded, Complete Answers or Correct Rejections?", elem_classes="section-heading")

                    with gr.Row(elem_id="summary-containers"):
                        with gr.Column(scale=1):
                            with gr.Group(elem_classes=["summary-card", "summary-card-a"]):
                                summary_a_display = gr.Textbox(
                                    label="Model A", 
                                    lines=10, 
                                    interactive=False, 
                                    show_copy_button=True, 
                                    autoscroll=False,
                                    elem_id="summary-a-display"
                                )
                        with gr.Column(scale=1):
                            with gr.Group(elem_classes=["summary-card", "summary-card-b"]):
                                summary_b_display = gr.Textbox(
                                    label="Model B", 
                                    lines=10, 
                                    interactive=False, 
                                    show_copy_button=True,
                                    autoscroll=False,
                                    elem_id="summary-b-display"
                                )
                    
                    
                    # NEW: Reference Answer Toggle (exactly like FAQ style)
                    with gr.Row(elem_id="reference-toggle-row"):
                        reference_toggle_btn = gr.Button("▶ Show Reference Answer", elem_classes=["faq-toggle-button"])
                    
                    # Reference Answer Content - initially hidden (exactly like FAQ)
                    with gr.Row(visible=False, elem_id="reference-content") as reference_content:
                        reference_answer_display = gr.Markdown("", elem_classes="faq-text")

                # Voting section - initially hidden
                with gr.Column(visible=False, elem_id="voting-section") as voting_section:
                    gr.HTML("<hr>")
                    gr.Markdown("### 🏅 Cast Your Vote", elem_classes="section-heading")
                    with gr.Row():
                        vote_button_a = gr.Button("⬅️ Summary A is Better", elem_classes=["vote-button"], interactive=False)
                        vote_button_tie = gr.Button("🤝 Tie / Equally Good", elem_classes=["vote-button"], interactive=False)
                        vote_button_b = gr.Button("➡️ Summary B is Better", elem_classes=["vote-button"], interactive=False)
                        vote_button_neither = gr.Button("❌ Neither is Good", elem_classes=["vote-button", "vote-button-neither"], interactive=False)

                with gr.Group(elem_classes=["feedback-section"], visible=False) as feedback_section:
                    feedback_checkboxes = gr.CheckboxGroup(label="Feedback (optional)", choices=[], interactive=False)
                
                # Submit button - initially hidden
                submit_button = gr.Button("Submit Your Vote", variant="primary", interactive=False, elem_id="submit-button", visible=False)

                with gr.Column(visible=False) as results_reveal_area:
                    gr.Markdown("---")
                    gr.Markdown("### ✅ Vote Submitted!", elem_classes="section-heading")
                     
                    with gr.Row():
                        with gr.Column(scale=1):
                            gr.Markdown("### Model A was:", elem_classes="section-heading")
                            model_a_reveal = gr.Markdown("", elem_classes="model-reveal model-a-reveal")
                        with gr.Column(scale=1):
                            gr.Markdown("### Model B was:", elem_classes="section-heading")
                            model_b_reveal = gr.Markdown("", elem_classes="model-reveal model-b-reveal")
                     
                    gr.HTML("<hr>")
                    
                    with gr.Row(elem_classes=["control-buttons"]):
                        try_another_btn = gr.Button("🔄 Try Another Question", elem_id="try-another-btn")

        with gr.TabItem("Leaderboard", id="leaderboard-tab"):
            gr.Markdown("# SLM RAG Leaderboard", elem_classes="orange-title")
            gr.HTML('View performance statistics for all models ranked by Elo rating. <br><br><a href="https://docs.google.com/forms/d/e/1FAIpQLSeUZoy43MlpK8-tJS4a6n5Q8PAKf-8Twdui5ybU18t0e2UuVA/viewform" class="form-link" target="_blank" rel="noopener noreferrer">Submit a new model request</a>')
            
            with gr.Group(elem_id="leaderboard-info"):
                gr.Markdown("""### About Elo Ratings
                
The Elo rating system provides a more accurate ranking than simple win rates:
- All models start at 1500 points
- Points are exchanged after each comparison based on the expected outcome
- Beating a stronger model earns more points than beating a weaker one
- The ± value shows the statistical confidence interval (95%)
""")
            
            results_table_display = gr.HTML(label="Model Performance")

    # FAQ toggle functionality with icon change
    faq_toggle_btn.click(
        fn=toggle_faq,
        inputs=[faq_expanded],
        outputs=[faq_expanded, faq_content, faq_toggle_btn]
    )
    
    # Context toggle functionality
    context_toggle_btn.click(
        fn=toggle_context_display,
        inputs=[current_example, show_full_context],
        outputs=[show_full_context, context_display, context_toggle_btn]
    )
    
    # NEW: Reference answer toggle functionality (exactly like FAQ)
    reference_toggle_btn.click(
        fn=toggle_reference_answer,
        inputs=[show_reference_answer, current_example],
        outputs=[show_reference_answer, reference_content, reference_toggle_btn, reference_answer_display]
    )

    # Initialize UI to empty state on load
    demo.load(
        fn=initialize_empty_app,
        inputs=[],
        outputs=[
            context_section,
            model_section,
            voting_section,
            submit_button
        ]
    )

    # Load leaderboard on start
    demo.load(
        fn=load_leaderboard,
        inputs=[],
        outputs=[results_table_display]
    )

    # Getting a new question
    random_question_btn.click(
        fn=show_loading_state,
        inputs=[],
        outputs=[
            summary_a_display, summary_b_display, 
            vote_button_a, vote_button_b, vote_button_tie, vote_button_neither,
            feedback_section, submit_button, results_reveal_area, random_question_btn,
            selected_winner
        ]
    ).then(
        fn=handle_new_example_click,
        inputs=[],
        outputs=[current_example]
    ).then(
        fn=update_ui_for_new_context,
        inputs=[current_example],
        outputs=[query_display, context_description, context_display, 
                context_toggle_btn, show_full_context]
    ).then(
        # NEW: Reset reference section when loading new question
        fn=reset_reference_section,
        inputs=[],
        outputs=[show_reference_answer, reference_content, reference_toggle_btn, reference_answer_display]
    ).then(
        # IMPORTANT: Explicitly hide FAQ here
        fn=hide_faq_section,
        inputs=[],
        outputs=[faq_container, faq_content]
    ).then(
        fn=show_all_after_loading,
        inputs=[],
        outputs=[
            context_section, 
            model_section, 
            voting_section,
            submit_button,
            random_question_btn
        ]
    ).then(
        fn=process_example,
        inputs=[current_example],
        outputs=[model_a_name, model_b_name, summary_a_text, summary_b_text,
                selected_winner, feedback_list, show_results_state, results_agg,
                summary_a_display, summary_b_display, vote_button_a, vote_button_b, 
                vote_button_tie, vote_button_neither, feedback_checkboxes, feedback_section, 
                submit_button, results_reveal_area, random_question_btn, main_interface_area]
    )

    # Try another question
    try_another_btn.click(
        fn=show_loading_state,
        inputs=[],
        outputs=[
            summary_a_display, summary_b_display, 
            vote_button_a, vote_button_b, vote_button_tie, vote_button_neither,
            feedback_section, submit_button, results_reveal_area, random_question_btn,
            selected_winner
        ]
    ).then(
        fn=handle_new_example_click,
        inputs=[],
        outputs=[current_example]
    ).then(
        fn=update_ui_for_new_context,
        inputs=[current_example],
        outputs=[query_display, context_description, context_display, 
                context_toggle_btn, show_full_context]
    ).then(
        # NEW: Reset reference section when trying another question
        fn=reset_reference_section,
        inputs=[],
        outputs=[show_reference_answer, reference_content, reference_toggle_btn, reference_answer_display]
    ).then(
        # IMPORTANT: Explicitly hide FAQ here too
        fn=hide_faq_section,
        inputs=[],
        outputs=[faq_container, faq_content]
    ).then(
        fn=show_all_after_loading,
        inputs=[],
        outputs=[
            context_section, 
            model_section, 
            voting_section,
            submit_button,
            random_question_btn
        ]
    ).then(
        fn=process_example,
        inputs=[current_example],
        outputs=[model_a_name, model_b_name, summary_a_text, summary_b_text,
                selected_winner, feedback_list, show_results_state, results_agg,
                summary_a_display, summary_b_display, vote_button_a, vote_button_b, 
                vote_button_tie, vote_button_neither, feedback_checkboxes, feedback_section, 
                submit_button, results_reveal_area, random_question_btn, main_interface_area]
    )

    # Vote button handling
    for btn, choice in zip(
        [vote_button_a, vote_button_b, vote_button_tie, vote_button_neither],
        ['left', 'right', 'tie', 'neither']
    ):
        btn.click(
            fn=lambda choice=choice: select_vote_improved(choice),
            inputs=None,
            outputs=[selected_winner, feedback_checkboxes, feedback_section, submit_button, 
                    vote_button_a, vote_button_b, vote_button_tie, vote_button_neither]
        )

    feedback_checkboxes.change(
        fn=update_feedback,
        inputs=[feedback_checkboxes],
        outputs=[feedback_list]
    )

    submit_button.click(
        fn=handle_vote_submission,
        inputs=[current_example, model_a_name, model_b_name, selected_winner, feedback_list, summary_a_text, summary_b_text, results_agg],
        outputs=[show_results_state, results_agg, vote_button_a, vote_button_b, 
                vote_button_tie, vote_button_neither, feedback_checkboxes,
                feedback_section, submit_button, results_reveal_area,
                random_question_btn, results_table_display, main_interface_area,
                context_toggle_btn, model_a_reveal, model_b_reveal]
    )
    
    tabs.select(
        fn=load_leaderboard,
        inputs=[],
        outputs=[results_table_display],
        api_name="refresh_leaderboard"
    )
    
    demo.unload(cleanup_on_disconnect)

if __name__ == "__main__":
    demo.launch(debug=True)