Spotify: The archive – the tech behind your 2025 wrapped highlights (engineering.atspotify.com)

Spotify has unveiled its enhanced Wrapped feature for 2025, dubbed Wrapped Archive, which offers users a deeper storytelling experience through their listening habits. By identifying up to five significant days in a user's year based on listening metrics, Spotify has leveraged a combination of heuristics and large language models (LLMs) to generate personalized reports that narrate each day's unique musical journey. The system's robust architecture was designed to handle over 1.4 billion individual reports efficiently, with a focus on data-driven storytelling that maintains emotional resonance and safety. The technical implementation included a distributed data pipeline and sophisticated prompt engineering to ensure consistent and meaningful outputs. Spotify used high-performance models during development but transitioned to a more efficient production model via a distillation pipeline, enabling the generation of reports at scale without compromising quality. Key innovations included Direct Preference Optimization for fine-tuning the LLM and careful data modeling to avoid concurrency issues. The initiative signifies a notable advancement for the AI/ML community, demonstrating effective integration of machine learning in real-time data applications and personalized user experiences at unprecedented scale.