Problem
A marketplace that adds grocery, retail and convenience next to restaurants has deep data on restaurant orders but little on the new categories. A customer is well known in one vertical and cold in the others. A catalog of tens to hundreds of thousands of SKUs spreads behavior too thinly for a standard recommender, and popularity baselines push the same head products to everyone. Relevance drops and long-tail items get buried.
Example from the article: a customer who keeps ordering Indian dishes (Butter Chicken, Vegetable Samosa, Naan) in restaurants is likely to want vegetable sides, chicken, naan and spices in grocery.
Solution in brief
An LLM reads each user's restaurant orders and search queries and returns affinities for categories of the grocery taxonomy, level by level, with a confidence score. Only confident categories are kept. They become sparse features of the existing multi-task ranker, so a user with no grocery history is personalized from the first session.
Questions this case answers
- How to transfer user preferences from a data-rich product category to a category with sparse interactions and a very large catalog
- How to generate stable, catalog-aligned user features with an LLM at the cost of a batch job
- How to recommend grocery and retail items to customers who have orders only in restaurants?
- How to turn raw orders and search queries into features a production ranker can use?
Tools and methods
- LLM taxonomy mapping (hierarchical RAG) — maps order and search history to L1-L4 product categories, top-down
- GPT-4o-mini — generates the affinities; GPT-4o is the judge for feature quality
- Prompt caching and just-in-time materialization — cut compute cost by about 80%
- Multi-task ranking model — jointly predicts click, add to cart and purchase, takes the affinities as input
- Human and LLM-as-a-judge evaluation — scores features before ranker training
Google or a one-time email code. No password.