Retrieve Hundreds, Rank Dozens

Your catalog has 50 million items. Your feed budget is 200 milliseconds, end to end. A rich ranking model that costs even 1ms per item would need ~14 GPU-hours per request. This is not an optimization problem — it's a physics problem, and every production recommender solves it the same way: a funnel.

• Candidate generation — cheap, index-driven, tuned for recall: millions of items → a few hundred candidates
• Ranking — expensive, feature-rich, tuned for precision: hundreds of candidates → a scored top ~20
• Re-ranking — slate-level rules and constraints applied to the final list

Each stage gets its own model, its own metric, and its own failure modes. Conflating them is how teams end up debugging 'the recommender' when the problem lives entirely in one stage.