app.py

import gradio as gr
from sentence_transformers import SentenceTransformer
import pickle, json
import numpy as np
from huggingface_hub import hf_hub_download

# ── Load model artifacts ──────────────────────────────────────────────
print("Loading classifier...")
CLASSIFIER_PATH = hf_hub_download(
    "AurelPx/hr-conversations-classifier",
    "setfit_classifier.pkl",
    repo_type="model"
)
LABEL_PATH = hf_hub_download(
    "AurelPx/hr-conversations-classifier",
    "setfit_label_config.json",
    repo_type="model"
)

encoder = SentenceTransformer('sentence-transformers/all-MiniLM-L6-v2')
with open(CLASSIFIER_PATH, 'rb') as f:
    classifier = pickle.load(f)
with open(LABEL_PATH) as f:
    config = json.load(f)

LABELS = config['label_names']
print(f"Loaded {len(LABELS)} labels: {LABELS}")

# ── Classification function ─────────────────────────────────────────────
def classify(text: str, threshold: float):
    if not text or not text.strip():
        return "Please enter a conversation.", ""
    
    emb = encoder.encode([text])
    proba = classifier.predict_proba(emb)
    
    # Build sorted probability list for all labels
    all_probs = []
    for i, p in enumerate(proba):
        prob = float(p[0][1])
        all_probs.append((LABELS[i], prob))
    all_probs.sort(key=lambda x: x[1], reverse=True)
    
    # Filter by threshold
    predicted = [(l, p) for l, p in all_probs if p >= threshold]
    
    if not predicted:
        pred_str = f"No labels above threshold {threshold}"
    else:
        pred_str = " | ".join([f"**{l}** ({p:.3f})" for l, p in predicted])
    
    probs_str = "\n".join([f"{l}: {p:.3f}" for l, p in all_probs])
    
    return pred_str, probs_str

# ── Gradio UI ──────────────────────────────────────────────────────────
with gr.Blocks(title="HR Conversations Classifier") as demo:
    gr.Markdown("""
    # 🏢 HR Conversations Multi-Label Classifier
    
    Classify HR support conversations into **20 topic labels**.
    
    | Metric | Score |
    |--------|-------|
    | **F1-micro (5-fold CV)** | **0.7962 ± 0.0098** |
    | **F1-macro (5-fold CV)** | **0.7721** |
    
    **Model**: SETFit (MiniLM-L6-v2 + Logistic Regression)  
    **Training data**: 5,100 conversations (5,000 synthetic + 100 real)
    """)
    
    with gr.Row():
        with gr.Column(scale=2):
            text_input = gr.Textbox(
                label="Conversation",
                placeholder="Paste an HR conversation here...\n\nExample:\nUSER: I haven't received my payslip for March yet. Could you please check what's going on?\nAGENT: Good morning. I've checked the payroll system and it appears your March payslip was generated on the 28th but there was a distribution delay.",
                lines=10
            )
            threshold = gr.Slider(0.0, 1.0, value=0.5, step=0.05, label="Classification Threshold")
            classify_btn = gr.Button("Classify", variant="primary")
        
        with gr.Column(scale=1):
            output_labels = gr.Textbox(label="Predicted Labels", lines=3)
            output_probs = gr.Textbox(label="All Probabilities (sorted)", lines=22)
    
    gr.Examples(
        examples=[
            ["USER: I haven't received my payslip for March yet. Could you please check what's going on?\nAGENT: Good morning. I've checked the payroll system and it appears your March payslip was generated on the 28th but there was a distribution delay. I've resent it to your registered email.", 0.5],
            ["USER: I need to take sick leave starting today. I woke up with a terrible flu and cant come in. Whats the proceedure?\nAGENT: I'm sorry to hear that. Please rest and take care of yourself. You need to submit a sick leave request in the HR portal and upload your medical certificate within 48 hours.", 0.5],
            ["USER: I would like to understand the rules around parental leave in France. My partner is expecting and I want to plan ahead.\nAGENT: Congratulations! Under French labor law, the second parent is entitled to 25 calendar days of paternity leave.", 0.5],
            ["USER: I received an email asking me to complete a GDPR refresher training. Is this mandatory?\nAGENT: Yes, the GDPR refresher is mandatory for all employees and must be completed annually.", 0.5],
            ["USER: I want to dispute my performance review. My manager gave me a rating that I believe is unfair and biased.\nAGENT: I'm sorry to hear that. You have the right to formally dispute your review. The first step is to submit a written appeal through the HR portal within 15 days.", 0.5],
        ],
        inputs=[text_input, threshold],
        label="Try these examples"
    )
    
    classify_btn.click(
        fn=classify,
        inputs=[text_input, threshold],
        outputs=[output_labels, output_probs]
    )

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