Yandex has launched Sona, a generative AI model designed to integrate multiple stages of a recommendation system into a single module. Unlike previous multi-stage algorithms, Sona autonomously analyzes user behavior, including selections, rejections, and revisits, to rank content recommendations without manual feature engineering.
The model replaces a complex cascade of dozens of algorithms that previously required individual training and tuning. By consolidating these tasks, Sona incorporates user history directly into the content selection process. The system was first deployed to manage music recommendations on smart speakers equipped with the Alice voice assistant.
Following a one-week implementation period, Sona fully replaced the legacy recommendation framework. Data from the rollout shows a 6.3% increase in listening time for recommended tracks and an 18% rise in repeat playback requests. The number of active listeners grew by 4.5%, which is 2.4 times higher than the growth observed after the previous system update.
