"""
Rebuilds the FAISS products index from the cleaned metadata.

The page_content for each product is built as:
    title + brand (if not Generic) + categories (filtered)

The categories are filtered to remove:
- "Clothing, Shoes & Jewelry" (present in 99.98%, zero discriminating value)
- Categories that are too generic (e.g. "Women", "Men" on their own)

This improves the relevance of the semantic search by preventing
irrelevant tokens from polluting the embedding.

Input:  artifacts/metadata_cleaned.json
Output: artifacts/faiss_products/

Estimated time: 3-5 minutes (100k embeddings).
"""
import json
import os
import sys
from langchain_huggingface import HuggingFaceEmbeddings
from langchain_community.vectorstores import FAISS
from langchain_core.documents import Document

PROJECT_ROOT = os.path.dirname(os.path.dirname(os.path.abspath(__file__)))
INPUT_FILE = os.path.join(PROJECT_ROOT, "artifacts", "metadata_cleaned.json")
INDEX_NAME = os.path.join(PROJECT_ROOT, "artifacts", "faiss_products")

# Categories to exclude from page_content (too generic, they pollute the embedding)
EXCLUDED_CATEGORIES = {
    "Clothing, Shoes & Jewelry",
    "Clothing, Shoes & Jewelry ",
}


def flatten_categories(categories):
    """Flattens categories (they can be strings or nested lists) and filters out the useless ones."""
    result = []
    for c in categories:
        if isinstance(c, str):
            if c not in EXCLUDED_CATEGORIES:
                result.append(c)
        elif isinstance(c, list):
            for sub in c:
                if isinstance(sub, str) and sub not in EXCLUDED_CATEGORIES:
                    result.append(sub)
    return result


def build_index():
    print("--- BUILDING SEARCH INDEX (FAISS) ---")

    if not os.path.exists(INPUT_FILE):
        print(f"Error: {INPUT_FILE} not found!")
        sys.exit(1)

    print("-> Loading cleaned products...")
    with open(INPUT_FILE, "r", encoding="utf-8") as f:
        products = json.load(f)

    print(f"-> Preparing documents ({len(products)} items)...")
    docs = []
    for p in products:
        # Title: always present
        parts = [p['title']]

        # Brand: only if not Generic (all Generic = zero value)
        brand = p.get('brand', 'Generic')
        if brand and brand != 'Generic':
            parts.append(brand)

        # Categories: filtered (removes "Clothing, Shoes & Jewelry")
        cats = flatten_categories(p.get('categories', []))
        if cats:
            parts.append(' '.join(cats))

        page_content = ' '.join(parts)
        meta = {
            "asin": p["asin"],
            "title": p["title"],
            "price": str(p["price"]),
            "brand": p["brand"],
            "image_url": p["image_url"],
        }
        docs.append(Document(page_content=page_content, metadata=meta))

    print("-> Computing embeddings and indexing (may take a few minutes)...")
    embeddings = HuggingFaceEmbeddings(model_name="paraphrase-multilingual-MiniLM-L12-v2")
    vectorstore = FAISS.from_documents(docs, embeddings)

    vectorstore.save_local(INDEX_NAME)
    print(f"Index saved to: {INDEX_NAME}")


if __name__ == "__main__":
    build_index()