import pandas as pd
import numpy as np
from sentence_transformers import SentenceTransformer
import os

# ==============================================================================
# SCRIPT: embedding.py
# DESCRIPTION: Converts text data into semantic vectors (embeddings).
# ==============================================================================

def generate_embeddings(input_path, output_metadata_path, output_vector_path):
    # 1. Data Loading
    if not os.path.exists(input_path):
        print(f"Error: {input_path} not found!")
        return

    print("Reading normalized POI data...")
    df = pd.read_json(input_path)

    # Validate that the necessary column exists for the model
    if 'enriched_description' not in df.columns:
        print("Error: 'enriched_description' column missing. Check the normalization step.")
        return

    # 2. Model Initialization (multilingual-e5-small)
    # This model is optimized for both Turkish and English semantic search.
    print("Loading model: multilingual-e5-small...")
    model = SentenceTransformer('intfloat/multilingual-e5-small')

    # 3. Text Preprocessing
    # The E5 model family requires the 'passage: ' prefix for documents to optimize retrieval.
    print("Preparing texts for vectorization...")
    texts = ["passage: " + str(text) for text in df['enriched_description']]

    # 4. Embedding Generation
    print(f"Computing embeddings for {len(texts)} locations... (This may take a moment)")
    # convert_to_numpy=True ensures the output is ready for binary storage and math operations.
    embeddings = model.encode(texts, show_progress_bar=True, convert_to_numpy=True)

    # 5. Modular Storage
    # A. Vector File (.npy): Stored in binary format for high-speed retrieval and low memory footprint.
    np.save(output_vector_path, embeddings)
    print(f"Vectors saved to: {output_vector_path}")

    # B. Metadata File (.json): Saves the original data without the heavy vector arrays.
    
    df_metadata = df.drop(columns=['embedding'], errors='ignore')
    df_metadata.to_json(output_metadata_path, orient="records", force_ascii=False, indent=4)
    print(f"Metadata saved to: {output_metadata_path}")

    print("\n Embedding process completed successfully!")

if __name__ == "__main__":
    
    # Get the absolute path of the directory where this script is located
    script_dir = os.path.dirname(os.path.abspath(__file__))
    
    # Navigate to the 'app' directory (one level up from 'scripts')
    app_dir = os.path.dirname(script_dir)
    
    # Define the data directory path
    data_dir = os.path.join(app_dir, "data")
    
    # Define file paths relative to the data directory
    INPUT_FILE = os.path.join(data_dir, "data_sightseeing_ready.json")
    META_OUTPUT = os.path.join(data_dir, "poi_metadata.json")
    VEC_OUTPUT = os.path.join(data_dir, "poi_vectors.npy")
    
    generate_embeddings(INPUT_FILE, META_OUTPUT, VEC_OUTPUT)