What else will they buy?
Load a real order and see what the shop suggests next, or open any of 3,500 products from a real wholesaler's catalogue and fill a basket yourself. All of it runs in your browser.
The demo needs JavaScript. The catalogue and the model are about 2.2 MB, downloaded once.
Real orders and prices from a UK online gift wholesaler, December 2009 to December 2011 (UCI Online Retail II, CC BY 4.0). Every suggestion comes from the orders placed before October 2011; the rates below are measured on the ten weeks after. Wholesale buyers take one design in several colours, so colour variants show up together.
33%Of 4,865 later real orders with one item left out, that item is among the shop's top ten suggestions, out of 3,529 products. Random guessing: 0.3%.
3%The same test for a bestseller list, what most shops show.
0.3%The same question put to a general AI model of the kind behind chatbots. It knows what a gift shop sells, not what these customers buy together.
How it works
DataTwo years of orders, 3,500 products, nothing about the customer beyond what was in the basket. Methods are fitted on orders before October 2011 and tested on the ten weeks after.
MethodsBestsellers, "bought together" from a linear item-to-item model (EASE) fitted on all orders, and an embedding model that learnt which products belong together. All three run on your device from a 2 MB JSON file. The embedding model scores 16% on the same test. The general model is Qwen 2.5 7B, an open LLM; given 100 products to pick from, it reaches 2.7%.
What you get"Frequently bought together" on every product page, "Complete your order" in the basket, the same lists in e-mails, from your own orders and refreshed nightly on any server. No GPU, no API bill, no customer data leaving your systems.
Code, data pipeline and the full benchmark: github.com/balticml/recommendation-engine. Want this on your shop? Get in touch