Problem
Yandex Lavka (grocery delivery) recommends items on about 10 app surfaces: home, product page, cart. Recommendations drive more than 20% of GMV, at about 1100 RPS at peak and a 300 ms deadline. The team ran a transformer over user history only as a daily batch job, so the ranker saw history that was 24 hours old or older. An offline test showed 2% higher NDCG with history that is 90 seconds old.
Example from the article: the feed is returned in chunks of 20 items, and add-to-cart conversion by position spikes at every multiple of 20. A user adds something from the first chunk, and the next chunk is requested with the updated cart.
Solution in brief
Lavka moved a late-interaction (two-tower) transformer to runtime. A history service turns recent actions into tensors, the inference service returns the user embedding, and the recommender fetches it asynchronously while collecting candidates. The gradient-boosting ranker uses the user-item dot products as features. After the ranker was retrained, cart revenue grew by 6-8%.
Questions this case answers
- How to add a real-time sequence model over a user's recent actions to a gradient-boosting ranker without retraining the whole ranker, while keeping training and serving preprocessing identical
- How to fetch a long user action history within a tight latency budget and feed its embedding into a ranking model
- How to add a real-time transformer over user history to a grocery recommender without retraining the whole ranker
- How to keep training and serving preprocessing identical for a sequence model
Tools and methods
- Two-tower transformer — user tower over an action sequence, item tower over content embeddings
- WordPiece tokenization — embeddings for new products without retraining
- Dynamic table (YTsaurus) — user history with fast reads and writes
- Shared preprocessing config — one code path for training and serving
- Gradient boosting ranker — consumes the transformer scores as features
Google or a one-time email code. No password.
