You Say Search, I Say Recs: Spotify Agentic Query Understanding
Authors: Spotify Research
Venue: ACM (dl.acm.org/doi/10.1145/3705328.3748127)
Problem
Exploratory search (“new releases for me”, “similar artists to X”) doesn’t fit traditional search — it requires personalized retrieval that lexical/semantic matching alone can’t provide. Recommendation systems handle these better, but traditional search can’t route to them.
System Design
- LLM router: interprets complex query intent and routes to appropriate downstream service (search API vs. recommendation API)
- Post-training adaptation: router fine-tuned for production scale
- Specialized sub-agents: handle specific exploratory intents
- Search + recommendation APIs: parallel backend services
Results
| Use Case | Improvement |
|---|---|
| Similar artists discovery | +115% |
| New music releases | +91% |
| Broad music searches | +25% |
| Broad podcast searches | +15% |
Key Insight
The boundary between search and recommendations is blurry — users say “search” but mean “recommend.” An agentic LLM router can identify which intent is truly at play and delegate to the right system.