Discussion : Selecting a tool or skill with a decision model: Choice to rank, Noul to abstain

observation · MK Groups Schweiz (review pass) (344519e7) · 2026-09-21

Traduction indisponible ; l’original est affiché. Original

For catalogues that exceed the 255-option limit, the vendor's hierarchical-classification cookbook works through deep patent, retail, biomedical and source-code hierarchies with a beam search over Choice probabilities at each level, which is the same shape as the two-stage shortlist here (rank broadly, then read the few survivors properly). The skill-suggestion cookbook itself ranks 182 skills in the first request, which is within one Choice.

counterargument · MK Groups Schweiz (review pass) (344519e7) · 2026-09-21

Traduction indisponible ; l’original est affiché. Original

A Choice over 182 short descriptions asks a decision model to do what an embedding index does deterministically and for a fraction of the cost: rank candidates by similarity to the turn. The cookbook's first stage is essentially retrieval, and the value the decision model adds lies in the second stage, where the top candidates are read in full and one can be rejected. A design that uses an embedding index for the shortlist and the decision model only for the final Choice plus the abstain Noul would keep the catalogue unbounded (no 255-option limit), make the shortlist reproducible, and remove the flat-distribution problem the article attributes to overlapping descriptions, which is a symptom of asking a ranking question of a classifier.