Spotify Engineering Reveals Tech Behind 2025 Wrapped Archive
Spotify Engineering has detailed the technical infrastructure powering the new 'Wrapped Archive' feature for its 2025 year-in-review campaign. This initiative moves beyond standard statistics by identifying up to five remarkable listening days for each user and generating personalized, narrative-driven reports using Large Language Models (LLMs). The system employs a priority-ordered set of heuristics to detect significant moments, such as biggest discovery days, nostalgic spikes, or unusual listening patterns. To manage the massive scale of generating 1.4 billion reports, Spotify implemented a distributed data pipeline and a sophisticated prompt engineering strategy. This involved splitting prompts into system and user layers to ensure data accuracy, brand consistency, and safety. Furthermore, the company utilized a model distillation pipeline, where a high-performance frontier model created a curated 'gold' dataset to fine-tune a smaller, more efficient production model. This approach allowed Spotify to deliver creative, emotionally resonant stories grounded in real listening data while maintaining economic feasibility and operational speed at a global scale.
Wire timeline
Spotify Engineering Reveals Tech Behind 2025 Wrapped Archive
Spotify Engineering has detailed the technical infrastructure powering the new 'Wrapped Archive' feature for its 2025 year-in-review campaign. This initiative moves beyond standard statistics by identifying up to five remarkable listening days for each user and generating personalized, narrative-driven reports using Large Language Models (LLMs). The system employs a priority-ordered set of heuristics to detect significant moments, such as biggest discovery days, nostalgic spikes, or unusual listening patterns. To manage the massive scale of generating 1.4 billion reports, Spotify implemented a distributed data pipeline and a sophisticated prompt engineering strategy. This involved splitting prompts into system and user layers to ensure data accuracy, brand consistency, and safety. Furthermore, the company utilized a model distillation pipeline, where a high-performance frontier model created a curated 'gold' dataset to fine-tune a smaller, more efficient production model. This approach allowed Spotify to deliver creative, emotionally resonant stories grounded in real listening data while maintaining economic feasibility and operational speed at a global scale.
Spotify Engineering