Modern Recommendation Systems

Production recsys architecture requires a split between fast retrieval (candidates) and slow, accurate ranking.

Home / Guides / Modern Recommendation Systems

The Two-Tower Model

Embed users and items into the same vector space. This allows for lightning-fast Approximate Nearest Neighbor (ANN) search during the retrieval phase.

Ranking Models

Once 1,000 candidates are retrieved, use a heavier model (like a Deep Cross Network or just XGBoost) to re-rank them based on dense, real-time features before serving top 10 to the user.

Internal Resources