Problem
Etsy has nearly 90M buyers and over 100 million listings. Classic search and recommendation systems miss nuanced tastes such as a preference for minimalist style or a current shopping mission like a fall sweater. A buyer-level profile could fix that, but asking an LLM to write one for every buyer was at first very expensive and took weeks per refresh.
Example from the article: the query "cool posters" is expanded to "cool posters + hippie|boho|vintage|nature" using the buyer's profile.
Solution in brief
Feed each buyer's recent sessions (searches, views, purchases, favorites, cart adds with listing text) to an LLM. It returns a structured profile: fandoms, styles, categories, an explanation and a confidence per interest. Cost and latency were cut with cheaper data access, a shorter history window, bigger batches and a smaller model. The profiles then drive query rewriting and refinement pills.
Questions this case answers
- How to turn raw behavioral logs of millions of users into structured interest profiles with an LLM at an affordable cost
- How to use a per-user interest profile to rewrite search queries and suggest refinements
- How to capture a buyer's styles and interests from browsing and purchase history with an LLM?
- How to make LLM-generated user profiles affordable for tens of millions of users?
Tools and methods
- LLM profile generation — turns session text into a JSON profile with confidence scores
- BigQuery and the internal feature store — source of events and listing titles and descriptions; sink for sessions and profiles
- Apache Airflow — schedules and staggers batches by user_id
- OpenAI API — the LLM endpoint, with a cached prompt
- Query rewriting — appends predicted interests to the user's query
- User-based collaborative filtering — gives new users an "inheritance profile" of similar users
Google or a one-time email code. No password.