Score Fusion

Combining several ranking signals into one number, and choosing what happens when they disagree — including how results from several sub-queries are merged into one context.

Why it matters in a memory layer

The weights are where a ranker is right or wrong, and they cannot be reasoned out: on this corpus the type weight had to be swept before 0.5 was defensible, and merging strategy mattered as much as scoring. Global top-k lets the better-matching half of a compound question take every slot; strict round-robin hands the weaker half padding while a needed fact waits. Guaranteeing each sub-question its best answer and filling the rest by score is what finally put all four required facts in a five-slot context.

Connections

Taught in: Hybrid Ranking

Used in: The Packing Problem