import gradio as gr
import torch
import joblib
import numpy as np
from transformers import BertTokenizer, BertModel

# ----------------- 1. Setup Device -----------------
# HF Spaces (Free) usually runs on CPU, but this keeps it robust
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
print(f"Using device: {device}")

# ----------------- 2. Load BERT -----------------
print("Loading BERT model...")
tokenizer = BertTokenizer.from_pretrained('bert-base-uncased')
bert_model = BertModel.from_pretrained('bert-base-uncased')
bert_model.to(device)
bert_model.eval()

# ----------------- 3. Load MLP + Scaler + LabelEncoder -----------------
# Ensure these files are uploaded to your HF Space Files tab!
print("Loading classification components...")
try:
    mlp = joblib.load("mlp_query_classifier.joblib")
    scaler = joblib.load("scaler_query_classifier.joblib")
    le = joblib.load("label_encoder_query_classifier.joblib")
    print("Loaded MLP, scaler, and label encoder.")
except FileNotFoundError as e:
    print(f"Error: {e}. Please make sure you uploaded the .joblib files to the Space.")

# ----------------- 4. Embedding Function -----------------
def get_bert_embeddings(text_list):
    inputs = tokenizer(
        text_list,
        padding=True,
        truncation=True,
        max_length=128,
        return_tensors="pt"
    ).to(device)

    with torch.no_grad():
        outputs = bert_model(**inputs)

    cls_embeddings = outputs.last_hidden_state[:, 0, :]
    return cls_embeddings.cpu().numpy()

# ----------------- 5. Prediction Function -----------------
def predict_new_query(text):
    # 1) BERT embedding
    embedding = get_bert_embeddings([text])
    
    # 2) scale with same scaler as training
    embedding_scaled = scaler.transform(embedding)
    
    # 3) MLP prediction -> class index
    prediction_index = mlp.predict(embedding_scaled)[0]
    
    # 4) map index back to string label
    label = le.inverse_transform([prediction_index])[0]
    
    # Optional: Get probability if your MLP supports it
    try:
        probs = mlp.predict_proba(embedding_scaled)[0]
        confidence = np.max(probs)
        return f"Label: {label} (Confidence: {confidence:.2f})"
    except:
        return f"Label: {label}"

# ----------------- 6. Launch Gradio Interface -----------------
# This creates the web UI
iface = gr.Interface(
    fn=predict_new_query,
    inputs=gr.Textbox(lines=2, placeholder="Enter your query here..."),
    outputs="text",
    title="BERT Query Classifier",
    description="Enter a text query to classify it using the custom BERT+MLP model."
)

iface.launch()