1
Curious builder0 XP earned · 300 to level 2
0 daysFinish a lesson to begin
Badge collection0 of 6 unlocked
51 small wins to finish your pathNext question →
How would you use LLMs to improve a recommendation system for a content platform?
30-second answerSay your answer out loud first, then reveal.
Where LLMs add value
| Use | How | Online/offline |
|---|---|---|
| Item enrichment | Extract topics, difficulty, entities from titles, descriptions, transcripts | Offline batch |
| Semantic item embeddings | Better similarity for content-based retrieval | Offline |
| Cold start | Recommend new items based on content features before clicks exist | Offline features, online serving |
| User interest summaries | Summarise long histories into interest profiles | Near-line, periodic |
| Explanations | "Recommended because you liked X" | Online, small model or cached |
| Conversational discovery | "Show me beginner-friendly RAG videos under 20 minutes" | Online, separate flow |
| Final re-ranking | LLM re-ranks top 20 for diversity or intent | Online, latency permitting |
Architecture sketch: candidate generation (two-tower embeddings, collaborative filtering, LLM-based content embeddings) → ranking model (gradient-boosted or deep model with behavioural + LLM-derived features) → re-ranking/business rules (diversity, freshness) → optional LLM explanation.
Evaluation: offline (recall@k, NDCG on held-out interactions) plus online A/B tests (watch time, retention, diversity, satisfaction surveys).
Pitfalls: LLM-generated features may encode popularity or demographic biases; cost of enrichment at catalogue scale (use batch); explanations must be faithful to the actual reasons.
Related
Every expert started right here.