← All models · DeBERTa zero-shot
use case · multi-model
Two independent ways to label an incoming message. MiniLM embeddings find the most similar labelled example you've seen before and borrow its label (nearest-neighbour). DeBERTa zero-shot scores your label set directly, using no examples at all. When the two methods agree, you can route automatically; when they disagree, flag it for a human. Both run locally.
Loads two models (~25 MB embeddings + ~70 MB zero-shot). Both auto-initialise when cached.
Method A · MiniLM k-NN over examples
Method B · DeBERTa zero-shot
Retrieval improves as your labelled memory grows (it's literally learning from examples); zero-shot works from day one with zero examples. Running both and comparing gives you a cheap, on-device confidence signal — a real pattern for bootstrapping a classifier before you have much labelled data.