Multi-Stage Ranking

A recommendation system with hundreds of millions of items can use embeddings to quickly narrow things down to a few thousand candidates in milliseconds.

But that’s only the first step.

The ranking model that decides the final order is much heavier. It uses hundreds of features and complex interactions, and it simply can’t score thousands of items within a tight latency budget.

Multi-stage ranking solves this by breaking the problem into steps. Start with a fast, lightweight model to filter the candidates, then pass a smaller set to progressively more powerful models. Each stage trims the list so the next one can afford to go deeper.