app.py

import gradio as gr
import torch
from transformers import CLIPProcessor, CLIPModel
import re

# Load FashionCLIP model
model_name = "patrickjohncyh/fashion-clip"
model = CLIPModel.from_pretrained(model_name)
processor = CLIPProcessor.from_pretrained(model_name)

# Price extraction regex
price_pattern = re.compile(r'(\bunder\b|\babove\b|\bbelow\b|\bbetween\b)?\s?(\d{1,5})\s?(AED|USD|EUR)?', re.IGNORECASE)

def get_text_embedding(text):
    """
    Converts input text into an embedding using FashionCLIP.
    """
    inputs = processor(text=[text], images=None, return_tensors="pt", padding=True)
    with torch.no_grad():
        text_embedding = model.get_text_features(**inputs)
    return text_embedding

def extract_attributes(query):
    """
    Extract structured fashion attributes dynamically using FashionCLIP.
    """
    structured_output = {"Brand": "Unknown", "Category": "Unknown", "Gender": "Unknown", "Price": "Unknown"}

    # Get embedding for the query
    query_embedding = get_text_embedding(query)

    # Compare with embeddings of common fashion attribute words (using FashionCLIP)
    reference_labels = ["Brand", "Category", "Gender", "Price"]
    reference_embeddings = get_text_embedding(reference_labels)

    # Compute cosine similarity to classify the type of query
    similarities = torch.nn.functional.cosine_similarity(query_embedding, reference_embeddings)
    best_match_index = similarities.argmax().item()

    # Assign type dynamically
    attribute_type = reference_labels[best_match_index]

    # Extract price dynamically
    price_match = price_pattern.search(query)
    if price_match:
        condition, amount, currency = price_match.groups()
        structured_output["Price"] = f"{condition.capitalize() if condition else ''} {amount} {currency if currency else 'AED'}".strip()

    # Extract brand & category dynamically using FashionCLIP similarity
    structured_output[attribute_type] = query  # Assigning full query text to matched attribute

    return structured_output

# Define Gradio UI
def parse_query(user_query):
    """
    Takes user query and returns structured attributes dynamically.
    """
    parsed_output = extract_attributes(user_query)
    return parsed_output  # Returns structured JSON

with gr.Blocks() as demo:
    gr.Markdown("# 🛍️ Fashion Query Parser using FashionCLIP")

    query_input = gr.Textbox(label="Enter your search query", placeholder="e.g., Gucci men’s perfume under 200AED")
    output_box = gr.JSON(label="Parsed Output")

    parse_button = gr.Button("Parse Query")
    parse_button.click(parse_query, inputs=[query_input], outputs=[output_box])

demo.launch()