public_eval_gradio.py

import gradio as gr
from gradio_modal import Modal
from huggingface_hub import hf_hub_download, list_repo_files
import os, csv, datetime, sys
import json
from utils import format_chat, append_to_sheet, read_sheet_to_df
import random

# Define the six evaluation criteria as a list of dictionaries.
criteria = [
    {
        "label": "Problem Resolution",
        "text": (
            "Problem Resolution: Did the model effectively solve the problem?",
            "1️⃣ Did Not Solve the Problem at All<br>2️⃣ Attempted to Solve but Largely Incorrect or Incomplete<br>3️⃣ Partially Solved the Problem, but with Limitations<br>4️⃣ Mostly Solved the Problem, with Minor Issues<br>5️⃣ Completely and Effectively Solved the Problem"
        )
    },
    {
        "label": "Helpfulness",
        "text": (
            "Helpfulness: Was the answer and reasoning provided helpful in addressing the question?",
            "1️⃣ Not Helpful at All<br>2️⃣ Slightly Helpful, but Largely Insufficient<br>3️⃣ Moderately Helpful, but Needs Improvement<br>4️⃣ Helpful and Mostly Clear, with Minor Issues<br>5️⃣ Extremely Helpful and Comprehensive"
        )
    },
    {
        "label": "Scientific Consensus",
        "text": (
            "Scientific and Clinical Consensus: Does the answer align with established scientific and clinical consensus?",
            "1️⃣ Completely Misaligned with Scientific Consensus<br>2️⃣ Partially Aligned but Contains Significant Inaccuracies or Misinterpretations<br>3️⃣ Generally Aligned but Lacks Strong Scientific Rigor or Clarity<br>4️⃣ Mostly Aligned with Scientific Consensus, with Minor Omissions or Uncertainties<br>5️⃣ Fully Aligned with Established Scientific and Clinical Consensus"
        )
    },
    {
        "label": "Accuracy",
        "text": (
            "Accuracy of Content: Is there any incorrect or irrelevant content in the answer and the reasoning content?",
            "1️⃣ Completely Inaccurate or Irrelevant<br>2️⃣ Mostly Inaccurate, with Some Relevant Elements<br>3️⃣ Partially Accurate, but Includes Some Errors or Omissions<br>4️⃣ Mostly Accurate, with Minor Issues or Unverified Claims<br>5️⃣ Completely Accurate and Relevant"
        )
    },
    {
        "label": "Completeness",
        "text": (
            "Completeness: Did the answer omit any essential content necessary for a comprehensive response?",
            "1️⃣ Severely Incomplete – Major Content Omissions<br>2️⃣ Largely Incomplete – Missing Key Elements<br>3️⃣ Somewhat Complete – Covers Basics but Lacks Depth<br>4️⃣ Mostly Complete – Minor Omissions or Gaps<br>5️⃣ Fully Complete – No Important Omissions"
        )
    },
]

criteria_for_comparison = [
    {
        "label": "Problem Resolution",
        "text": (
            "Problem Resolution: Did the model effectively solve the problem?<br>"
        )
    },
    {
        "label": "Helpfulness",
        "text": (
            "Helpfulness: Was the answer and reasoning provided helpful in addressing the question?<br>"
        )
    },
    {
        "label": "Scientific Consensus",
        "text": (
            "Scientific and Clinical Consensus: Does the answer align with established scientific and clinical consensus?<br>"
        )
    },
    {
        "label": "Accuracy",
        "text": (
            "Accuracy of Content: Is there any incorrect or irrelevant content in the answer and the reasoning content?<br>"
        )
    },
    {
        "label": "Completeness",
        "text": (
            "Completeness: Did the answer omit any essential content necessary for a comprehensive response?<br>"
        )
    },
]

mapping = {   #for pairwise mapping between model comparison selections
    "👈 Model A": "A",
    "👉 Model B": "B",
    "🤝 Tie": "tie",
    "👎 Neither model did well": "neither"
}

#Prepare data
REPO_ID  = "RichardZhu52/TxAgent_human_eval"
CROWDSOURCING_DATA_DIRECTORY = "crowdsourcing_eval_data_0430"
TXAGENT_RESULTS_SHEET_BASE_NAME = "TxAgent_Human_Eval_Results_CROWDSOURCED"
DISEASE_SPECIALTY_MAP_FILENAME = "disease_specialty_map.json"

def get_evaluator_questions(email, disease_map_data, user_all_specs, all_files, evaluator_directory):
    relevant_diseases = []
    for disease, specs in disease_map_data.items():
        disease_specs = set(specs.get('specialties', []))
        disease_subspecs = set(specs.get('subspecialties', []))
        
        # Check for intersection
        if user_all_specs.intersection(disease_specs) or user_all_specs.intersection(disease_subspecs):
            relevant_diseases.append(disease)

    # Filter to only the files in that directory
    evaluator_files = [f for f in all_files if f.startswith(f"{evaluator_directory}/")]
    data_by_filename = {}
    for remote_path in evaluator_files:
        local_path = hf_hub_download(
            repo_id=REPO_ID,
            repo_type="dataset",
            revision="main", #fetches the most recent version of the dataset each time this command is called
            filename=remote_path,
            # force_download=True,
        )
        with open(local_path, "r") as f:
            model_name_key = os.path.basename(remote_path).replace('.json', '')
            data_by_filename[model_name_key] = json.load(f)

    # Filter questions based on relevant diseases derived from user specialties
    evaluator_question_ids = []
    relevant_diseases_lower = {disease.lower() for disease in relevant_diseases} # Convert relevant diseases to lowercase set for efficient lookup

    # Assuming 'txagent' data is representative for question IDs and associated diseases
    if 'txagent' in data_by_filename:
        for entry in data_by_filename['txagent']:
            question_id = entry.get("question_ID")
            question_diseases = entry.get("disease", []) # Get diseases list, default to empty if missing
            if question_id is not None and question_diseases:
                # Convert question diseases to lowercase and check for intersection
                question_diseases_lower = {disease.lower() for disease in question_diseases if isinstance(disease, str)}
                if question_diseases_lower.intersection(relevant_diseases_lower):
                    evaluator_question_ids.append(question_id)

    # Handle case where no relevant questions are found based on specialty
    if not evaluator_question_ids:
        return gr.update(visible=True), gr.update(visible=False), None, "No questions found matching your selected specialties. Please adjust your selections or contact the study administrator.", gr.Chatbot(), gr.Chatbot(), gr.HTML(),gr.State(),gr.update(visible=False),""

    #FINALLY, MAKE SURE THEY DIDNT ALREADY FILL IT OUT. Must go through every tuple of (question_ID, TxAgent, other model) where other model could be any of the other files in data_by_filename
    model_names = [key for key in data_by_filename.keys() if key != 'txagent']
    # evaluator_question_ids = question_map.get(email).get('question_ids')
    full_question_ids_list = []
    for other_model_name in model_names:
        for q_id in evaluator_question_ids:
            full_question_ids_list.append((q_id, other_model_name))

    results_df = read_sheet_to_df(custom_sheet_name=str(TXAGENT_RESULTS_SHEET_BASE_NAME))
    if (results_df is not None) and (not results_df.empty):
        # collect all (question_ID, other_model) pairs already seen
        matched_pairs = set()
        for _, row in results_df.iterrows():
            # Only consider rows submitted by the current evaluator
            print(f"Evaluator ID: {row['Email']}")
            if row["Email"] == email:
                q = row["Question ID"]
                # pick whichever response isn’t 'txagent'
                a, b = row["ResponseA_Model"], row["ResponseB_Model"]
                if a == "txagent" and b != "txagent":
                    matched_pairs.add((q, b))
                elif b == "txagent" and a != "txagent":
                    matched_pairs.add((q, a))

        # filter out any tuple whose (q_id, other_model) was already matched
        full_question_ids_list = [
            (q_id, other_model)
            for (q_id, other_model) in full_question_ids_list
            if (q_id, other_model) not in matched_pairs
        ]
        print(f"Filtered question IDs: {full_question_ids_list}")
        print(f"Length of filtered question IDs: {len(full_question_ids_list)}")

    
    return full_question_ids_list, data_by_filename


def go_to_eval_progress_modal(name, email, specialty_dd, subspecialty_dd, years_exp_radio, exp_explanation_tb):

    # ADDED: Validate that name and email are non-empty before proceeding
    if not name or not email:
        return gr.update(visible=True), gr.update(visible=False), None, "Please fill out all the fields.", gr.Chatbot(), gr.Chatbot(), gr.HTML(),gr.State(),gr.update(visible=False), ""

    # Combine user's selected specialties and subspecialties into a set for efficient lookup
    # Ensure inputs are lists, even if None or single strings are passed (though Dropdown with multiselect=True should return lists)
    user_specialties = set(specialty_dd if isinstance(specialty_dd, list) else ([specialty_dd] if specialty_dd else []))
    user_subspecialties = set(subspecialty_dd if isinstance(subspecialty_dd, list) else ([subspecialty_dd] if subspecialty_dd else []))
    user_all_specs = user_specialties.union(user_subspecialties)

    #retrieve data from HF
    evaluator_directory = CROWDSOURCING_DATA_DIRECTORY
    if evaluator_directory is None:
        return gr.update(visible=True), gr.update(visible=False), None, "Invalid Evaluator ID, please try again.", gr.Chatbot(), gr.Chatbot(), gr.HTML(),gr.State(),gr.update(visible=False),"" 
    all_files = list_repo_files(
        repo_id=REPO_ID,
        repo_type="dataset",
        revision="main",
    )

    disease_specialty_map = hf_hub_download(
        repo_id=REPO_ID,
        filename=DISEASE_SPECIALTY_MAP_FILENAME,
        repo_type="dataset",
        revision="main",
    )

    with open(disease_specialty_map, 'r') as f:
        disease_map_data = json.load(f)

    full_question_ids_list, data_by_filename = get_evaluator_questions(email, disease_map_data, user_all_specs, all_files, evaluator_directory)

    if len(full_question_ids_list) == 0:
        return gr.update(visible=True), gr.update(visible=False), None, "Based on your Evaluator ID, you have evaluated all questions assigned to you. You may exit the page; we will follow-up if we require anything else from you. Thank you!", gr.Chatbot(), gr.Chatbot(), gr.HTML(),gr.State(),gr.update(visible=False),""

    full_question_ids_list = sorted(full_question_ids_list, key=lambda x: str(x[0])+str(x[1]))
    #selected question is the first element
    q_id, other_model_name = full_question_ids_list[0]

    #Constructing question_for_eval, the question to evaluate this round
    txagent_matched_entry = next(
        (entry for entry in data_by_filename['txagent'] if entry.get("question_ID") == q_id),
        None
    )
    other_model_matched_entry = next(
        (entry for entry in data_by_filename[other_model_name] if entry.get("question_ID") == q_id),
        None
    )

    models_list = [
        {
            "model": "txagent",
            "reasoning_trace": txagent_matched_entry.get("solution")
        },
        {
            "model": other_model_name,
            "reasoning_trace": other_model_matched_entry.get("solution")
        }
    ]
    random.shuffle(models_list)

    question_for_eval = {
        "question": txagent_matched_entry.get("question"),
        "question_ID": q_id,
        "models": models_list,
    }

    #update user_info
    user_info = (name, email, specialty_dd, subspecialty_dd, years_exp_radio, exp_explanation_tb, q_id)
    chat_A_value = format_chat(question_for_eval['models'][0]['reasoning_trace'])
    chat_B_value = format_chat(question_for_eval['models'][1]['reasoning_trace'])
    prompt_text = question_for_eval['question']

    # Construct the question-specific elements of the pairwise rating page (page 1)
    page1_prompt = gr.HTML(f'<div style="background-color: #FFEFD5; border: 2px solid #FF8C00; padding: 10px; border-radius: 5px; color: black;"><strong style="color: black;">Prompt:</strong> {prompt_text}</div>')
    chat_a = gr.Chatbot(
                    value=chat_A_value,
                    type="messages",
                    height=400,
                    label="Model A Response",
                    show_copy_button=False,
                    show_label=True,
                    render_markdown=True,  # Required for markdown/HTML support in messages
                    avatar_images=None,    # Optional: omit user/assistant icons
                    rtl=False
                )
    chat_b = gr.Chatbot(
                    value=chat_B_value,
                    type="messages",
                    height=400,
                    label="Model B Response",
                    show_copy_button=False,
                    show_label=True,
                    render_markdown=True,  # Required for markdown/HTML support in messages
                    avatar_images=None,    # Optional: omit user/assistant icons
                    rtl=False
                )
    return gr.update(visible=True), gr.update(visible=False), user_info,"", chat_a, chat_b, page1_prompt, question_for_eval, gr.update(visible=True), f"You have {len(full_question_ids_list)} question(s) remaining to evaluate."

#goes to page 1 from confirmation modal that tells users how many questions they have left to evaluate
def go_to_page1():
    """
    Shows page 1
    """

    # Return updates to hide modal, hide page 0, show page 1, populate page 1, and set final state
    updates = [
        gr.update(visible=False),
        gr.update(visible=False),
        gr.update(visible=True),
    ]
    return updates


# Callback to transition from Page 1 to Page 2.
def go_to_page2(data_subset_state,*pairwise_values):
    # pairwise_values is a tuple of values from each radio input.
    criteria_count = len(criteria_for_comparison)
    pairwise_list = list(pairwise_values[:criteria_count])
    comparison_reasons_list = list(pairwise_values[criteria_count:])

    #gradio components to display previous page results on next page
    pairwise_results_for_display = [gr.Markdown(f"***As a reminder, your pairwise comparison answer for this criterion was: {pairwise_list[i]}. Your answer choices will be restricted based on your comparison answer, but you may go back and change the comparison answer if you wish.***") for i in range(len(criteria))]

    if any(answer is None for answer in pairwise_list):
        return gr.update(visible=True), gr.update(visible=False), None, None, "Error: Please select an option for every pairwise comparison.", gr.Chatbot(), gr.Chatbot(), gr.HTML(), *pairwise_results_for_display

    chat_A_value = format_chat(data_subset_state['models'][0]['reasoning_trace'])
    chat_B_value = format_chat(data_subset_state['models'][1]['reasoning_trace'])
    prompt_text = data_subset_state['question']

    # Construct the question-specific elements of the rating page (page 2)
    chat_A_rating = gr.Chatbot(
                    value=chat_A_value,
                    type="messages",
                    height=400,
                    label="Model A Response",
                    show_copy_button=False,
                    render_markdown=True
                )

    chat_B_rating = gr.Chatbot(
                    value=chat_B_value,
                    type="messages",
                    height=400,
                    label="Model B Response",
                    show_copy_button=False,
                    render_markdown=True
                )

    page2_prompt = gr.HTML(f'<div style="background-color: #FFEFD5; border: 2px solid #FF8C00; padding: 10px; border-radius: 5px; color: black;"><strong style="color: black;">Prompt:</strong> {prompt_text}</div>')
    
    return gr.update(visible=False), gr.update(visible=True), pairwise_list, comparison_reasons_list, "", chat_A_rating, chat_B_rating, page2_prompt, *pairwise_results_for_display


# Callback to store scores for Response A.
def store_A_scores(*args):
    # Unpack the arguments: first half are scores, second half are checkboxes.
    num = len(args) // 2
    scores = list(args[:num])
    unquals = list(args[num:])
    return scores, unquals

# Callback to transition from Page 2 to Page 3.
def go_to_page3():
    return gr.update(visible=False), gr.update(visible=True)

# Updated validation callback that ignores criteria with 'Unable to Judge'
def validate_ratings(pairwise_choices, *args):
    num_criteria = len(criteria)
    ratings_A_list = list(args[:num_criteria])
    ratings_B_list = list(args[num_criteria:])
    if any(r is None for r in ratings_A_list) or any(r is None for r in ratings_B_list):
        return "Error: Please provide ratings for both responses for every criterion.", "Error: Please provide ratings for both responses for every criterion."
    error_msgs = []
    for i, choice in enumerate(pairwise_choices):
        score_a = ratings_A_list[i]
        score_b = ratings_B_list[i]
        # Skip criteria if either rating is "Unable to Judge"
        if score_a == "Unable to Judge" or score_b == "Unable to Judge":
            continue
        # Convert string scores to integers for comparison.
        score_a = int(score_a)
        score_b = int(score_b)
        if choice == "👈 Model A" and score_a < score_b:
            error_msgs.append(f"Criterion {i+1} ({criteria[i]['label']}): You selected A as better but scored A lower than B.")
        elif choice == "👉 Model B" and score_b < score_a:
            error_msgs.append(f"Criterion {i+1} ({criteria[i]['label']}): You selected B as better but scored B lower than A.")
        elif choice == "🤝 Tie" and score_a != score_b:
            error_msgs.append(f"Criterion {i+1} ({criteria[i]['label']}): You selected Tie but scored A and B differently.")

    if error_msgs:
        err_str = "\n".join(error_msgs)
        return err_str, err_str
    else:
        return "No errors in responses; feel free to submit!", "No errors in responses; feel free to submit!"

# # Additional callback to handle submission results.
def toggle_slider(is_unqualified):
    # When the checkbox is checked (True), set interactive to False to disable the slider.
    return gr.update(interactive=not is_unqualified)

centered_col_css = """
#centered-column {
    margin-left: auto;
    margin-right: auto;
    max-width: 650px; /* Adjust this width as desired */
    width: 100%;
}
"""
with gr.Blocks(css=centered_col_css) as demo:
    # States to save information between pages.
    user_info_state = gr.State()
    pairwise_state = gr.State()
    scores_A_state = gr.State()
    comparison_reasons = gr.State()
    unqualified_A_state = gr.State()
    data_subset_state = gr.State()

    # Load specialty data
    specialties_path = os.path.join("specialties_preprocessing", "specialties.json")
    subspecialties_path = os.path.join("specialties_preprocessing", "subspecialties.json")

    try:
        with open(specialties_path, 'r') as f:
            specialties_list = json.load(f)
        with open(subspecialties_path, 'r') as f:
            subspecialties_list = json.load(f)
    except FileNotFoundError:
        print(f"Error: Could not find specialty files at {specialties_path} or {subspecialties_path}. Please ensure these files exist.")
        # Provide default empty lists or handle the error as appropriate
        specialties_list = ["Error loading specialties"]
        subspecialties_list = ["Error loading subspecialties"]
    except json.JSONDecodeError:
        print(f"Error: Could not parse JSON from specialty files.")
        specialties_list = ["Error parsing specialties"]
        subspecialties_list = ["Error parsing subspecialties"]

    # Page 0: Welcome / Informational page.
    with gr.Column(visible=True, elem_id="page0") as page0:
        gr.Markdown("## Welcome to the TxAgent User Study!")
        gr.Markdown("Please read the following instructions and then enter your information to begin:")
        # Existing informational markdown...
        gr.Markdown("""Thank you for your interest in TxAgent!

    TxAgent is a first-of-its-kind AI model developed in the Zitnik Lab at Harvard Medical School. It leverages multi-step reasoning and real-time biomedical knowledge retrieval from a toolbox of 211 tools to analyze drug interactions, contraindications, and patient-specific treatment strategies. It is designed to provide personalized treatment recommendations across a wide range of diseases, including rare diseases.

    TxAgent evaluates how drugs interact at molecular, pharmacokinetic, and clinical levels, identifies contraindications based on patient comorbidities and concurrent medications, and tailors treatment strategies to individual patient characteristics, including age, genetic factors, and disease progression. It achieves 92.1% accuracy in open-ended drug reasoning tasks, surpassing GPT-4o by up to 25.8% and outperforming DeepSeek-R1 (671B) in structured multi-step reasoning.

    TxAgent's toolbox, ToolUniverse, consolidates 211 tools from trusted sources, including all US FDA-approved drugs since 1939 and validated clinical insights from Open Targets. By integrating multi-step inference, real-time knowledge grounding, and tool-assisted decision-making, TxAgent ensures that treatment recommendations align with established clinical guidelines and real-world evidence, reducing the risk of adverse events and improving therapeutic decision-making.

    We are currently conducting a user study with physicians, rare disease experts, and others with relevant medical background to assess TxAgent's performance on personalized therapeutic reasoning across multiple criteria, including helpfulness, clinical consensus, and scientific accuracy. Please note that each session requires a minimum commitment of 5-10 minutes to complete one question. If you wish to evaluate multiple questions, you may do so; you will never be asked to re-evaluate questions you have already seen. By clicking 'Next' below, you will start the study, with your progress saved after submitting each question. When evaluating a question, you will be asked to compare the responses of two different models to the question and then rate each model's response on a scale of 1-5. You may use the Back and Next buttons at the bottom of each page to edit any of your responses before submitting. If you have any other questions or concerns, please contact us directly. Thank you for your participation!""")
        gr.Markdown("## Please enter your information to get a question to evaluate. You will only be able to evaluate one question at a time. You must submit and return to this page to evaluate the next question.")
        name = gr.Textbox(label="Name")
        email = gr.Textbox(label="Email (Please use the same email every time you log onto this portal, as we use your email to prevent showing duplicate questions.)")
        specialty_dd = gr.Dropdown(choices=specialties_list, label="Primary Medical Specialty (select one; go to https://www.abms.org/member-boards/specialty-subspecialty-certificates/ for categorization)", multiselect=True)
        subspecialty_dd = gr.Dropdown(choices=subspecialties_list, label="Subspecialty (if applicable, select one; go to https://www.abms.org/member-boards/specialty-subspecialty-certificates/ for categorization)", multiselect=True)
        years_exp_radio = gr.Radio(
            choices=["0-2 years", "3-5 years", "6-10 years", "11-20 years", "20+ years", "Not Applicable"],
            label="How many years have you been involved in clinical and/or research activities related to your biomedical area of expertise?"
        )
        exp_explanation_tb = gr.Textbox(label="Please briefly explain your expertise/experience relevant to evaluating AI for clinical decision support (optional)")

        page0_error_box = gr.Markdown("")
        next_btn_0 = gr.Button("Next")


    with Modal(visible=False) as eval_progress_modal:
        eval_progress_text = gr.Markdown("You have X questions remaining.")
        eval_progress_proceed_btn = gr.Button("OK, proceed to question evaluation")

    # Page 1: Pairwise Comparison.
    with gr.Column(visible=False) as page1:
        gr.Markdown("## Part 1/2: Pairwise Comparison") #Make the number controlled by question indexing!
        page1_prompt = gr.HTML()
        with gr.Row():
            # ADDED: Use gr.Chatbot to display the scrollable chat window for Response A.
            with gr.Column():
                gr.Markdown("**Model A Response:**")  # Already bold label.
                chat_a = gr.Chatbot(
                    value=[], # Placeholder for chat history
                    type="messages",
                    height=400,
                    label="Model A Response",
                    show_copy_button=False,
                    show_label=True,
                    render_markdown=True,  # Required for markdown/HTML support in messages
                    avatar_images=None,    # Optional: omit user/assistant icons
                    rtl=False
                )
            # ADDED: Use gr.Chatbot to display the scrollable chat window for Response B.
            with gr.Column():
                gr.Markdown("**Model B Response:**")
                chat_b = gr.Chatbot(
                    value=[],
                    type="messages",
                    height=400,
                    label="Model B Response",
                    show_copy_button=False,
                    show_label=True,
                    render_markdown=True,  # Required for markdown/HTML support in messages
                    avatar_images=None,    # Optional: omit user/assistant icons
                    rtl=False
                )
        gr.Markdown("<br><br>")
        gr.Markdown("### For each criterion, select which response did better:")
        comparison_reasons_inputs = []  # ADDED: list to store the free-text inputs
        pairwise_inputs = []
        for crit in criteria_for_comparison:
            with gr.Row():
                gr.Markdown(crit['text'])
                radio = gr.Radio(
                    choices=[
                        "👈 Model A",  # A
                        "👉 Model B",  # B
                        "🤝 Tie",      # tie
                        "👎 Neither model did well"  # neither
                    ],
                    label="Which is better?"
                )
                pairwise_inputs.append(radio)
            # ADDED: free text under each comparison
            text_input = gr.Textbox(label=f"Reasons for your selection (optional)")
            comparison_reasons_inputs.append(text_input)
        
        page1_error_box = gr.Markdown("")  # ADDED: display validation errors
        with gr.Row():
            back_btn_0 = gr.Button("Back")
            home_btn_1 = gr.Button("Home / FAQ")  # ADDED: FAQ button on page1
            next_btn_1 = gr.Button("Next: Rate Responses")
    
    # Page 2: Combined Rating Page for both responses.
    with gr.Column(visible=False) as page2:
        gr.Markdown("## Part 2/2: Rate Model Responses")
        # ### EDIT: Show a highlighted prompt as on previous pages.
        page2_prompt = gr.HTML()
        # ### EDIT: Display both responses side-by-side using Chatbot windows.
        with gr.Row():
            with gr.Column():
                gr.Markdown("**Model A Response:**")
                chat_a_rating = gr.Chatbot(
                    value=[],
                    type="messages",
                    height=400,
                    label="Model A Response",
                    show_copy_button=False,
                    render_markdown=True
                )
            with gr.Column():
                gr.Markdown("**Model B Response:**")
                chat_b_rating = gr.Chatbot(
                    value=[],
                    type="messages",
                    height=400,
                    label="Model B Response",
                    show_copy_button=False,
                    render_markdown=True
                )
        gr.Markdown("<br><br>")
        gr.Markdown("### For each criterion, select your ratings for each model response:")
        # ### EDIT: For each criterion, create a row with two multiple-choice sets (left: Response A, right: Response B) separated by a border.
        ratings_A = []  # to store the radio components for response A
        ratings_B = []  # to store the radio components for response B
        
        def restrict_choices(pairwise_list, index, score_a, score_b):
            """
            Returns (update_for_A, update_for_B).
            Enforces rating constraints based on the pairwise choice for the given criterion index.
            """
            # Get the specific pairwise choice for this criterion using the index
            # Add error handling in case the state/list is not ready or index is wrong
            if not pairwise_list or index >= len(pairwise_list):
                pairwise_choice = None
            else:
                pairwise_choice = pairwise_list[index]

            base = ["1","2","3","4","5","Unable to Judge"]
            # Default: no restrictions unless explicitly set
            upd_A = gr.update(choices=base)
            upd_B = gr.update(choices=base)

            # Skip if no meaningful pairwise choice or either score is "Unable to Judge"
            if pairwise_choice is None or pairwise_choice == "👎 Neither model did well" or (score_a is None and score_b is None):
                # If one score is UJ but the other isn't, AND it's a Tie, we might still want to restrict the non-UJ one later?
                # For now, keep it simple: if either is UJ or choice is Neither/None, don't restrict.
                return upd_A, upd_B

            # Helper to parse int safely
            def to_int(x):
                try: return int(x)
                except (ValueError, TypeError): return None

            a_int = to_int(score_a)
            b_int = to_int(score_b)

            # --- Apply Restrictions ---
            if pairwise_choice == "👈 Model A":
                # B must be ≤ A (if A is numeric)
                if a_int is not None: #it is None if unable to judge
                    allowed_b_choices = [str(i) for i in range(1, a_int + 1)] + ["Unable to Judge"]
                    current_b = score_b if score_b in allowed_b_choices else None # Keep current valid choice
                    upd_B = gr.update(choices=allowed_b_choices, value=current_b)
                # If A is UJ or non-numeric, B is unrestricted by this rule
                # else: upd_B remains gr.update(choices=base)
                if b_int is not None:
                    # A must be >= B (if B is numeric)
                    allowed_a_choices = [str(i) for i in range(b_int, 6)] + ["Unable to Judge"]
                    current_a = score_a if score_a in allowed_a_choices else None # Keep current valid choice
                    upd_A = gr.update(choices=allowed_a_choices, value=current_a)
                # If B is UJ or non-numeric, A is unrestricted by this rule
                # else: upd_A remains gr.update(choices=base)

            elif pairwise_choice == "👉 Model B":
                # A must be ≤ B (if B is numeric)
                if b_int is not None:
                    allowed_a_choices = [str(i) for i in range(1, b_int + 1)] + ["Unable to Judge"]
                    current_a = score_a if score_a in allowed_a_choices else None # Keep current valid choice
                    upd_A = gr.update(choices=allowed_a_choices, value=current_a)
                # If B is UJ or non-numeric, A is unrestricted by this rule
                # else: upd_A remains gr.update(choices=base)
                if a_int is not None:
                    # B must be >= A (if A is numeric)
                    allowed_b_choices = [str(i) for i in range(a_int, 6)] + ["Unable to Judge"]
                    current_b = score_b if score_b in allowed_b_choices else None # Keep current valid choice
                    upd_B = gr.update(choices=allowed_b_choices, value=current_b)
                # If A is UJ or non-numeric, B is unrestricted by this rule
                # else: upd_B remains gr.update(choices=base)

            elif pairwise_choice == "🤝 Tie":
                # If both are numeric, they must match. Enforce based on the one that *just changed*.
                # If one changes to numeric, force the other (if also numeric) to match.
                # If one changes to UJ, the other is unrestricted.
                if a_int is not None:
                    upd_B = gr.update(choices=[score_a])
                elif score_a == "Unable to Judge":
                    upd_B = gr.update(choices=["Unable to Judge"])
                if b_int is not None:
                    upd_A = gr.update(choices=[score_b])
                elif score_b == "Unable to Judge":
                    upd_A = gr.update(choices=["Unable to Judge"])

            return upd_A, upd_B

        def clear_selection():
            return None, None
        
        pairwise_results_for_display = [gr.Markdown(render=False) for _ in range(len(criteria))]
        indices_for_change = []
        for i, crit in enumerate(criteria):
            index_component = gr.Number(value=i, visible=False, interactive=False)
            indices_for_change.append(index_component)

            with gr.Column(elem_id="centered-column"):
                gr.Markdown(f'<div style="text-align: left;">{crit["text"][0]}</div>')
                gr.Markdown(f'<div style="text-align: left;">{crit["text"][1]}</div>')
                pairwise_results_for_display[i].render()
            with gr.Row():
                with gr.Column(scale=1):
                    rating_a = gr.Radio(choices=["1", "2", "3", "4", "5", "Unable to Judge"],
                                            label=f"Score for Response A - {crit['label']}",
                                            interactive=True)
                with gr.Column(scale=1):
                    rating_b = gr.Radio(choices=["1", "2", "3", "4", "5", "Unable to Judge"],
                                            label=f"Score for Response B - {crit['label']}",
                                            interactive=True)
            with gr.Row():
                clear_btn = gr.Button("Clear Selection")
                clear_btn.click(fn=clear_selection, outputs=[rating_a,rating_b])

                # wire each to re‐restrict the other on change
                rating_a.change(
                    fn=restrict_choices,
                    inputs=[ pairwise_state, index_component, rating_a, rating_b ],
                    outputs=[ rating_a, rating_b ]
                )
                rating_b.change(
                    fn=restrict_choices,
                    inputs=[ pairwise_state, index_component, rating_a, rating_b ],
                    outputs=[ rating_a, rating_b ]
                )
                ratings_A.append(rating_a)
                ratings_B.append(rating_b)
        with gr.Row():
            back_btn_2 = gr.Button("Back")
            home_btn_2 = gr.Button("Home / FAQ")  # ADDED: FAQ button on page1
            submit_btn = gr.Button("Submit (Note: Once submitted, you cannot edit your responses)", elem_id="submit_btn")
        result_text = gr.Textbox(label="Validation Result")
    
    # Final Page: Thank you message.
    with gr.Column(visible=False, elem_id="final_page") as final_page:
        gr.Markdown("## You have no questions left to evaluate. Thank you for your participation!")
        eval_again_btn = gr.Button("Evaluate Another Question")
    
    # Error Modal: For displaying validation errors.
    with Modal("Error", visible=False, elem_id="error_modal") as error_modal:
        error_message_box = gr.Markdown()
        ok_btn = gr.Button("OK")
        # Clicking OK hides the modal.
        ok_btn.click(lambda: gr.update(visible=False), None, error_modal)
    
    # Confirmation Modal: Ask for final submission confirmation.
    with Modal("Confirm Submission", visible=False, elem_id="confirm_modal") as confirm_modal:
        gr.Markdown("Are you sure you want to submit? Once submitted, you cannot edit your responses.")
        with gr.Row():
            yes_btn = gr.Button("Yes, please submit")
            cancel_btn = gr.Button("Cancel")
    
    # --- Define Callback Functions for Confirmation Flow ---
    def build_row_dict(data_subset_state, user_info, pairwise, comparisons_reasons, *args):
        num_criteria = len(criteria)
        ratings_A_vals = list(args[:num_criteria])
        ratings_B_vals = list(args[num_criteria:])

        prompt_text = data_subset_state['question']
        response_A_model = data_subset_state['models'][0]['model']
        response_B_model = data_subset_state['models'][1]['model']

        timestamp = datetime.datetime.now().isoformat()
        row = {
            "Timestamp": timestamp,
            "Name": user_info[0],
            "Email": user_info[1],
            "Specialty": str(user_info[2]),
            "Subspecialty": str(user_info[3]),
            "Years of Experience": user_info[4],
            "Experience Explanation": user_info[5],
            "Question ID": user_info[6],
            "Prompt": prompt_text,
            "ResponseA_Model": response_A_model,
            "ResponseB_Model": response_B_model,
        }

        pairwise = [mapping.get(val, val) for val in pairwise]
        for i, crit in enumerate(criteria):
            label = crit['label']
            row[f"Criterion_{label} Comparison: Which is Better?"] = pairwise[i]
            row[f"Criterion_{label} Comments"] = comparisons_reasons[i]
            row[f"ScoreA_{label}"] = ratings_A_vals[i]
            row[f"ScoreB_{label}"] = ratings_B_vals[i]

        return row

    # def final_submit(data_subset_state, user_info, pairwise, comparisons_reasons, *args):

    #     row_dict = build_row_dict(data_subset_state, user_info, pairwise, comparisons_reasons, *args)
    #     append_to_sheet(user_data=None, custom_row_dict=row_dict, custom_sheet_name=str(TXAGENT_RESULTS_SHEET_BASE_NAME), add_header_when_create_sheet=True)

    #     return gr.update(visible=False), gr.update(visible=True), gr.update(visible=False)

    def final_submit(data_subset_state, user_info, pairwise, comparisons_reasons, *args):
        # --- Part 1: Submit the current results (Existing Logic) ---
        row_dict = build_row_dict(data_subset_state, user_info, pairwise, comparisons_reasons, *args)
        append_to_sheet(user_data=None, custom_row_dict=row_dict, custom_sheet_name=str(TXAGENT_RESULTS_SHEET_BASE_NAME), add_header_when_create_sheet=True)

        # --- Part 2: Recalculate remaining questions (Existing Logic + Modified Error Handling) ---
        # try:

        # --- Re-fetch data and filter questions (Same logic as before) ---
        evaluator_directory = CROWDSOURCING_DATA_DIRECTORY
        all_files = list_repo_files(repo_id=REPO_ID, repo_type="dataset", revision="main")
        disease_specialty_map = hf_hub_download(repo_id=REPO_ID, filename=DISEASE_SPECIALTY_MAP_FILENAME, repo_type="dataset", revision="main")
        with open(disease_specialty_map, 'r') as f: disease_map_data = json.load(f)

        _, email, specialty, subspecialty, _, _, _ = user_info
        user_specialties = set(specialty if isinstance(specialty, list) else ([specialty] if specialty else []))
        user_subspecialties = set(subspecialty if isinstance(subspecialty, list) else ([subspecialty] if subspecialty else []))
        user_all_specs = user_specialties.union(user_subspecialties)

        full_question_ids_list, data_by_filename = get_evaluator_questions(email, disease_map_data, user_all_specs, all_files, evaluator_directory)
        remaining_count = len(full_question_ids_list)

        # --- Part 3: Determine UI updates based on remaining count ---
        if remaining_count == 0:
            # Success with NO remaining questions
            return (
                gr.update(visible=False),  # page0 (Hide)
                gr.update(visible=False),  # page2 (Hide)
                gr.update(visible=False),  # confirm_modal
                gr.update(visible=False), 
                "",                       
                gr.update(visible=True),   # final_page (Show)
                "",
                None,
                None,
                None,
                None
            )
        
        full_question_ids_list = sorted(full_question_ids_list, key=lambda x: str(x[0])+str(x[1]))
        #selected question is the first element
        q_id, other_model_name = full_question_ids_list[0]

        #Constructing question_for_eval, the question to evaluate this round
        txagent_matched_entry = next(
            (entry for entry in data_by_filename['txagent'] if entry.get("question_ID") == q_id),
            None
        )
        other_model_matched_entry = next(
            (entry for entry in data_by_filename[other_model_name] if entry.get("question_ID") == q_id),
            None
        )

        models_list = [
            {
                "model": "txagent",
                "reasoning_trace": txagent_matched_entry.get("solution")
            },
            {
                "model": other_model_name,
                "reasoning_trace": other_model_matched_entry.get("solution")
            }
        ]
        random.shuffle(models_list)

        question_for_eval = {
            "question": txagent_matched_entry.get("question"),
            "question_ID": q_id,
            "models": models_list,
        }

        chat_A_value = format_chat(question_for_eval['models'][0]['reasoning_trace'])
        chat_B_value = format_chat(question_for_eval['models'][1]['reasoning_trace'])
        prompt_text = question_for_eval['question']

        # Construct the question-specific elements of the pairwise rating page (page 1)
        page1_prompt = gr.HTML(f'<div style="background-color: #FFEFD5; border: 2px solid #FF8C00; padding: 10px; border-radius: 5px; color: black;"><strong style="color: black;">Prompt:</strong> {prompt_text}</div>')
        chat_a = gr.Chatbot(
                value=chat_A_value,
                type="messages",
                height=400,
                label="Model A Response",
                show_copy_button=False,
                show_label=True,
                render_markdown=True,  # Required for markdown/HTML support in messages
                avatar_images=None,    # Optional: omit user/assistant icons
                rtl=False
            )
        chat_b = gr.Chatbot(
                value=chat_B_value,
                type="messages",
                height=400,
                label="Model B Response",
                show_copy_button=False,
                show_label=True,
                render_markdown=True,  # Required for markdown/HTML support in messages
                avatar_images=None,    # Optional: omit user/assistant icons
                rtl=False
            )

        # Success with remaining questions
        return (
            gr.update(visible=False),  # page0 (Hide)
            gr.update(visible=False),  # page2 (Hide)
            gr.update(visible=False),  # confirm_modal (Hide)
            gr.update(visible=True),   # eval_progress_modal (Show)
            f"Submission successful! You have {remaining_count} question(s) remaining to evaluate. You may exit the page and return later if you wish.", # eval_progress_text
            gr.update(visible=False),  # final_page (Hide)
            "",         
            chat_a,
            chat_b,
            page1_prompt,
            question_for_eval)

        # except Exception as e:
        #     error_message = f"Your submission was saved, but an error occurred while checking for remaining questions: {e}. Please try starting the process again by entering your details. If the problem persists, contact the administrator."
        #     print(f"Error during recalculation in final_submit: {e}") # Keep logging for debugging
        #     # *** MODIFIED RETURN ***: Error during recalculation
        #     return (
        #         gr.update(visible=True),   # page0 (Show) - Send user back to start
        #         gr.update(visible=False),  # page2 (Hide)
        #         gr.update(visible=False),  # confirm_modal (Hide)
        #         gr.update(visible=False),  # eval_progress_modal (Hide)
        #         "",                        # eval_progress_text (Clear)
        #         gr.update(visible=False),  # final_page (Hide)
        #         error_message              # page0_error_box (Update with error)
        #     )
    
    def cancel_submission():
        # Cancel final submission: just hide the confirmation modal.
        return gr.update(visible=False)
    
    def reset_everything_except_user_info():

        # 3) Reset all pairwise radios & textboxes
        reset_pairwise_radios = [gr.update(value=None) for i in range(len(criteria))]
        reset_pairwise_reasoning_texts  = [gr.update(value=None) for i in range(len(criteria))]

        # 4) Reset all rating radios
        reset_ratings_A = [gr.update(value=None) for i in range(len(criteria))]
        reset_ratings_B = [gr.update(value=None) for i in range(len(criteria))]

        return (
            # pages
            gr.update(visible=True),  # page0
            gr.update(visible=False), # final_page

            # states
            # gr.update(value=None),  # user_info_state
            gr.update(value=None),  # pairwise_state
            gr.update(value=None),  # scores_A_state
            gr.update(value=None),  # comparison_reasons
            gr.update(value=None),  # unqualified_A_state
            # gr.update(value=None),  # data_subset_state

            #page0 elements that need to be reset
            gr.update(value=""),  #page0_error_box 

            # page1 elements that need to be reset
            # gr.update(value=""),  #page1_prompt
            # gr.update(value=[]),  #chat_a
            # gr.update(value=[]),  #chat_b
            gr.update(value=""),  #page1_error_box

            # page2 elements that need to be reset
            gr.update(value=""),  #page2_prompt
            gr.update(value=[]),  #chat_a_rating
            gr.update(value=[]),  #chat_b_rating
            gr.update(value=""),  #result_text

            #lists of gradio elements that need to be unrolled
            *reset_pairwise_radios,
            *reset_pairwise_reasoning_texts,
            *reset_ratings_A,
            *reset_ratings_B
        )
    
    # --- Define Transitions Between Pages ---

    # Transition from Page 0 (Welcome) to Page 1.
    next_btn_0.click(
        fn=go_to_eval_progress_modal,
        inputs=[name, email, specialty_dd, subspecialty_dd, years_exp_radio, exp_explanation_tb],
        outputs=[page0, page1, user_info_state, page0_error_box, chat_a, chat_b, page1_prompt, data_subset_state,eval_progress_modal,eval_progress_text],
        scroll_to_output=True
    )

    eval_progress_proceed_btn.click(
        fn=go_to_page1,
        inputs=None,
        outputs=[eval_progress_modal, page0, page1],
        scroll_to_output=True
    )

    # FAQ buttons simply bring back to page0
    home_btn_1.click(lambda: (gr.update(visible=True), gr.update(visible=False)), None, [page0, page1])
    home_btn_2.click(lambda: (gr.update(visible=True), gr.update(visible=False)), None, [page0, page2])
    
    # Transition from Page 1 to Page 0 (Back button).
    back_btn_0.click(
        fn=lambda: (gr.update(visible=True), gr.update(visible=False)),
        inputs=None,
        outputs=[page0, page1]
    )
    
    # Transition from Page 1 (Pairwise) to the combined Rating Page (Page 2).
    next_btn_1.click(
        fn=go_to_page2,  # ### EDIT: Rename or update the function to simply pass the pairwise inputs if needed.
        inputs=[data_subset_state,*pairwise_inputs,*comparison_reasons_inputs],
        outputs=[page1, page2, pairwise_state, comparison_reasons, page1_error_box, chat_a_rating, chat_b_rating, page2_prompt, *pairwise_results_for_display],
        scroll_to_output=True
    )
    
    # Transition from Rating Page (Page 2) back to Pairwise page.
    back_btn_2.click(
        fn=lambda: (gr.update(visible=True), gr.update(visible=False)),
        inputs=None,
        outputs=[page1, page2],
        scroll_to_output=True
    )
    
    # --- Submission: Validate the Ratings and then Process the Result ---
    def process_result(result):
        # If validation passed, show the confirmation modal and proceed.
        if result == "No errors in responses; feel free to submit!":
            return (
                gr.update(),  # Show page 3
                gr.update(),  # Hide final page
                gr.update(visible=True),  # Show confirmation modal
                gr.update(visible=False),  # Hide error modal
                gr.update(value="")  # EDIT: Clear the error_message_box
            )
        else:
            # If validation fails, show the error modal and display the error in error_message_box.
            return (
                gr.update(),             # Keep page3 as is
                gr.update(),             # Keep final page unchanged
                gr.update(visible=False), # Hide confirmation modal
                gr.update(visible=True),  # Show error modal
                gr.update(value=result)   # EDIT: Update error_message_box with the validation error
            )

    # ### EDIT: Update the submission callback to use the new radio inputs.
    submit_btn.click(
        fn=validate_ratings,
        inputs=[pairwise_state, *ratings_A, *ratings_B],
        outputs=[error_message_box, result_text]
    ).then(
        fn=process_result,
        inputs=error_message_box,
        outputs=[page2, final_page, confirm_modal, error_modal, error_message_box]
    )
    
    # Finalize submission if user confirms.
    # yes_btn.click(
    #     fn=final_submit,
    #     inputs=[data_subset_state, user_info_state, pairwise_state, comparison_reasons, *ratings_A, *ratings_B],
    #     outputs=[page2, final_page, confirm_modal]
    # )
    question_submission_event = yes_btn.click(
        fn=final_submit,
        inputs=[data_subset_state, user_info_state, pairwise_state, comparison_reasons, *ratings_A, *ratings_B],
        outputs=[
            page0,                 # Controlled by final_submit return value 1
            page2,                 # Controlled by final_submit return value 2
            confirm_modal,         # Controlled by final_submit return value 3
            eval_progress_modal,   # Controlled by final_submit return value 4
            eval_progress_text,    # Controlled by final_submit return value 5
            final_page,            # Controlled by final_submit return value 6
            page0_error_box,
            chat_a,
            chat_b,
            page1_prompt,
            data_subset_state
        ],
        scroll_to_output=True
    )
    
    # Cancel final submission.
    cancel_btn.click(
        fn=cancel_submission,
        inputs=None,
        outputs=confirm_modal
    )

    # Reset everything and evaluate another question button
    question_submission_event.then(
        fn=reset_everything_except_user_info,
        inputs=[],
        outputs=[
            # pages
            page0,
            final_page,

            # states
            # user_info_state,
            pairwise_state,
            scores_A_state,
            comparison_reasons,
            unqualified_A_state,
            # data_subset_state,

            #page0 elements that need to be reset
            page0_error_box,

            # # page1 elements that need to be reset
            # page1_prompt,
            # chat_a,
            # chat_b,
            page1_error_box,

            # page2 elements that need to be reset
            page2_prompt,
            chat_a_rating,
            chat_b_rating,
            result_text,

            #lists of gradio elements that need to be unrolled
            *pairwise_inputs,
            *comparison_reasons_inputs,
            *ratings_A,
            *ratings_B
        ]
    )

demo.launch(share=True)