Can I vibe code Wiser?

Build a lightweight product recommendation engine as a Shopify App using Node.js, Remix, Supabase (PostgreSQL with pgvector), and Shopify App Extensions.

1. DATA MODEL:
- Store catalog: `products` (id, shopify_id, title, handle, image_url, price, embedding vector(1536))
- Order graph: `order_items` (order_id, product_id)
- Item co-occurrence cache: `product_pairs` (product_a_id, product_b_id, weight integer)
- Interactions: `events` (session_id, product_id, event_type ['view', 'add_to_cart'], created_at)

2. CORE FUNCTIONALITY:
- Webhooks: Listen to `orders/create` to populate `order_items` and increment co-occurrence scores in `product_pairs` for all product combinations in the order.
- Sync Endpoint: Catalog webhooks (`products/create`, `products/update`) that send product descriptions to OpenAI Text-Embedding-3-Small to save vector embeddings.
- Recommendation API:
  a) `GET /api/recommendations/frequently-bought?product_id=X`: Returns top 4 product_ids from `product_pairs` ordered by weight descending.
  b) `GET /api/recommendations/related?product_id=X`: Returns top 4 similar products using pgvector cosine distance `<->` on product embeddings.
  c) `GET /api/recommendations/cart`: Accepts array of product_ids in cart, calculates aggregate vector mean, and returns closest non-cart items.
- Storefront Frontend: Provide a Liquid Theme App Extension containing a Web Component widget (`<product-recs type="frequently-bought" product-id="{{ product.id }}">`) that fetches from the API and renders a responsive grid with direct 'Add to Cart' buttons.

3. FAILURE MODES & EDGE CASES:
- Fallback Strategy: If `product_pairs` or vector similarity returns <4 items (cold start), pad results with top-selling global products.
- Out of Stock Handling: Filter recommendation queries against `inventory_quantity > 0` fetched from Shopify API or cache.
- Latency Optimization: Cache API responses in Redis or Cloudflare KV for 1 hour per product_id.

4. OUT OF SCOPE:
- Do not build complex A/B testing visual editors.
- Do not build external email embedding generators.
- Exclude manual merchant rule override dashboards in MVP.

$ each button prefixes agent-specific run instructions · build your own product, never copy proprietary code, trademarks or designs