jev classify

Assign labels from a set you define: one label, several labels, or a path through a hierarchy.

jev classify [text] (--labels <list> | --labels-json <ref> | --taxonomy <ref>) [--multi] [--other] [options]

When to use

Label issues, route tickets, tag content, sort documents into a catalog. Prefer it over ask --choice when the task is "pick a bucket": it adds an other escape, confidence gating, multi-label, and hierarchies.

Modes

Mode	Flag	Returns
Single (default)	--labels a,b:description,c	One label, confidence, full probabilities, action (auto or review)
Multi	--multi	One probability per label; applied lists those at or above --threshold
Taxonomy	--taxonomy @tree.json	A path decided one level per request, with per-level confidence

Options

Option	Meaning	Default
[text]	Text, @file, or -	stdin
-l, --labels <list>	Comma-separated labels, optionally label:description. Escape a literal comma as \,.
--labels-json <ref>	["a","b"], [{"label","description"}], or {"label": "description"}
-t, --taxonomy <ref>	Nested JSON, e.g. {"hardware": {"laptop": null, "phone": null}, "software": ["os", "app"]}
--multi	One yes/no probability per label instead of picking one	off
--other	Add an other option so the model can say nothing fits (single and taxonomy)	off
-i, --instructions <text>	Replace the default question
--min-confidence <p>	Below this confidence the result is review	0.6
--threshold <p>	Multi: probability at or above which a label applies	0.5
--state-json	Treat the input as JSON
--fail-on <list>	Exit 2 on review, other, or unlabeled (multi: nothing applied)	none

Example

jev classify "Login fails after the latest update" \
  -l "bug:defect in an existing feature,feature:new capability,question" --other

bug  conf 0.91  auto
[bug 0.93, question 0.04, feature 0.02, other 0.01]

Tips

• Put your domain rules in the label descriptions, not in the label names.
• Always add --other unless every input is known to fit a label.
• A taxonomy costs one request per level. Leaf lists (["os", "app"]) are shorthand for {"os": null, "app": null}.

Patterns: classification using confidence, hierarchical classification.