Scores here are hand-authored features, not a real model. Real rerankers are cross-encoders, late-interaction models, LLM judges, or learned rankers. Latency and quality depend on model, hardware, batching, and corpus. Tap any candidate row for its feature breakdown.