Zero-shot classification
Label text with categories you choose on the spot, without training anything.
Last updated
Analogy
A switch statement that understands English. Hand it the cases at runtime.
Code example
const result = await classify("Fix null pointer crash on login page", [ "bug", "feature", "refactor",]);// result.labels[0] === "bug"In an agent
Perfect for a router step or labelling PRs: fast and free compared with an LLM call.
Questions people ask
What is zero-shot classification?
Zero-shot classification labels text with categories you choose at the moment you call it, without training anything. You pass the text and a list of labels such as bug, feature and refactor, and the model scores each label. It works like a switch statement that understands plain English.
Can a small classifier replace an LLM for routing?
Often, yes. Routing and labelling only need to pick one of a few categories, and a small zero-shot model does that fast and for free on your own server. Keep the large LLM for the open-ended reasoning that comes after the route is chosen. Check accuracy on real examples first.