Two-Tower Retrieval and Deep Ranking

~50 min read

Designing the candidate generation and ranking stages: two-tower architecture, multi-source retrieval, multi-task ranking model, and feature engineering.

Two-Tower Retrieval and Deep Ranking is a Pro topic

Sign in, then upgrade to Pro or Power to unlock this topic and the full AI Engineering curriculum.

Key points

  • Two-tower models pre-compute video embeddings offline so only the user embedding is computed at query time — this is what enables < 10ms retrieval from 10M videos
  • Multi-source candidate generation (2-tower + co-watch + content-based + trending) is essential — no single source has both personalized and fresh coverage
  • Multi-task ranking (predict watch time + like + share - abandon as separate heads) allows business-goal tuning via weight adjustment without retraining the model