Modern music streaming platforms constantly balance exploiting familiar content with encouraging exploration of new items. Users often want to discover fresh music but hesitate to replace known favorites. Providing natural‑language explanatory reasons can lower this psychological barrier. Large language models (LLMs) excel at generating nuanced explanations, yet their real‑time inference cost and computational load make direct online deployment impractical.
We present a decoupled recommendation architecture that scales exploration without adding latency. LLM inference is moved to an asynchronous offline pipeline, where personalized pools of undiscovered artists are pre‑computed for each user together with tailored rationales. Offline jobs run in batch with distributed scheduling, persisting candidate pools and explanations to an online cache; the online stage performs only lightweight retrieval and ranking.
Large‑scale online A/B experiments show that adding LLM‑generated explanations significantly boosts click‑through and playback duration for new artists, raising overall exploration by roughly 12% and overall engagement by about 5%. The results validate the hypothesis that explanations reduce the trust barrier and increase exploration.
The design demonstrates how, in high‑throughput, low‑latency production environments, a layered approach of offline LLM reasoning and online retrieval can bring explainable recommendations to scale.
Review