Cascade Models

Multi-stage ranking pushes every candidate through each stage, gradually narrowing the set. Cascade models work differently. They handle easy cases with lightweight models and only escalate to more complex ones when confidence is low.

Most inputs are easy, and you don't need a large model to settle them. A spam filter can block known scam templates with rules or sender reputation. A content moderation system can flag content matching a known banned perceptual hash without ever running a transformer or vision model.

By reserving expensive models for ambiguous cases, cascades reduce average cost while preserving accuracy on the cases where accuracy is hardest.