app.py

import gradio as gr
from transformers import pipeline
import json

# Load model
print("Loading model...")
classifier = pipeline(
    "text-classification",
    model="JohnLicode/ethics-review-deberta",
    device="cpu"  # Use CPU for free tier
)
print("Model loaded!")

def classify_ethics(text: str, guideline_id: str = "", guideline_name: str = ""):
    """Classify single text for ethics guideline compliance."""
    
    # Format input like training data
    if guideline_id and guideline_name:
        input_text = f"Guideline {guideline_id} {guideline_name}: {text}"
    else:
        input_text = text
    
    # Truncate if too long
    input_text = input_text[:1500]
    
    # Get prediction
    result = classifier(input_text)[0]
    
    # Map labels
    label = result['label']
    if label == "LABEL_0":
        label = "ADDRESSED"
    elif label == "LABEL_1":
        label = "NEEDS_REVISION"
    
    return {
        "label": label,
        "score": round(result['score'], 4),
        "input_preview": input_text[:100] + "..."
    }


def classify_batch(batch_json: str):
    """
    Classify multiple texts in a single API call for better performance.
    
    Input: JSON string with format:
    [
        {"text": "...", "guideline_id": "1.1", "guideline_name": "Objectives"},
        {"text": "...", "guideline_id": "3.2", "guideline_name": "Privacy"},
        ...
    ]
    
    Output: JSON string with results for each input.
    """
    try:
        items = json.loads(batch_json)
    except json.JSONDecodeError as e:
        return json.dumps({"error": f"Invalid JSON: {str(e)}"})
    
    if not isinstance(items, list):
        return json.dumps({"error": "Input must be a JSON array"})
    
    if len(items) > 50:
        return json.dumps({"error": "Maximum 50 items per batch"})
    
    # Prepare all inputs
    formatted_inputs = []
    for item in items:
        text = item.get("text", "")
        g_id = item.get("guideline_id", "")
        g_name = item.get("guideline_name", "")
        
        if g_id and g_name:
            input_text = f"Guideline {g_id} {g_name}: {text}"
        else:
            input_text = text
        
        formatted_inputs.append(input_text[:1500])
    
    # Run batch inference (much faster than individual calls)
    predictions = classifier(formatted_inputs)
    
    # Format results
    results = []
    for pred in predictions:
        label = pred['label']
        if label == "LABEL_0":
            label = "ADDRESSED"
        elif label == "LABEL_1":
            label = "NEEDS_REVISION"
        
        results.append({
            "label": label,
            "score": round(pred['score'], 4)
        })
    
    return json.dumps(results)


# Create Gradio interface with both single and batch endpoints
with gr.Blocks(title="Ethics Review Classifier") as demo:
    gr.Markdown("# Ethics Review Classifier")
    gr.Markdown("Classify research proposal text against ethics guidelines. Returns ADDRESSED or NEEDS_REVISION.")
    
    with gr.Tab("Single Classification"):
        with gr.Row():
            with gr.Column():
                text_input = gr.Textbox(label="Text to Analyze", lines=5, placeholder="Enter the text from research proposal...")
                id_input = gr.Textbox(label="Guideline ID (optional)", placeholder="e.g., 1.1")
                name_input = gr.Textbox(label="Guideline Name (optional)", placeholder="e.g., Objectives")
                single_btn = gr.Button("Classify", variant="primary")
            with gr.Column():
                single_output = gr.JSON(label="Result")
        
        single_btn.click(classify_ethics, inputs=[text_input, id_input, name_input], outputs=single_output)
        
        gr.Examples(
            examples=[
                ["The general objective is to develop an AI ethics review system. Specific objectives: 1) Create scanning module 2) Implement matching.", "1.1", "Objectives"],
                ["All participant data will be encrypted using AES-256 and stored securely.", "3.2", "Privacy and confidentiality"],
                ["The study explores innovative approaches.", "1.7", "Sampling design and size"],
            ],
            inputs=[text_input, id_input, name_input],
        )
    
    with gr.Tab("Batch Classification (Fast)"):
        gr.Markdown("**For API users:** Send up to 50 items in one request for faster processing.")
        batch_input = gr.Textbox(
            label="Batch Input (JSON Array)", 
            lines=10, 
            placeholder='[{"text": "...", "guideline_id": "1.1", "guideline_name": "Objectives"}, ...]'
        )
        batch_btn = gr.Button("Classify Batch", variant="primary")
        batch_output = gr.Textbox(label="Batch Results (JSON)", lines=10)
        
        batch_btn.click(classify_batch, inputs=[batch_input], outputs=batch_output)

# Launch with API enabled
demo.launch()