Skip to content
Agentic WorkFlow

Classify Text

Classify Text

Section titled “Classify Text”

Sort text into your own labels — any words you like, no training — running locally in your browser with a zero-shot model (e.g. DeBERTa zero-shot on transformers.js). Private, offline, no cost.

What it does

Section titled “What it does”

Scores the text against every label you list using a zero-shot (NLI) model, then returns the best label and a score for each one. If even the best label scores below the Confidence threshold, label is null, so you can route “none of the above” separately.

When to use it

Section titled “When to use it”
  • Route incoming messages, tickets, or emails by topic or intent (billing, bug report, feature request, other).
  • Tag scraped content or notes with categories you invent on the spot.
  • Anywhere you’d use the Text Classifier agent but want it free, private, and offline without wiring a model dependency.

Inputs and settings

Section titled “Inputs and settings”
SettingNotes
TextThe text to classify.
LabelsComma-separated list of labels, e.g. billing, bug report, other. An expression may return a list or a comma/newline-separated string (e.g. {{ $json.labels }}). Duplicates are ignored (case-insensitive).
ModelAn installed zero-shot (NLI) model. Only zero-shot models are listed here.
Allow several labelsOff by default: labels compete for one answer (scores sum to 1). On: each label is scored independently, so several can be high at once.
Confidence thresholdBelow this score the top label is null (default 0.3).

Outputs

Section titled “Outputs”

Returns { label, score, labels }:

  • label — the best label, or null when its score is below the threshold.
  • score — the top label’s score (reported even when label is null).
  • labels — every label as { label, score }, highest first.

Example workflow

Section titled “Example workflow”

Feed an email body into Text, set Labels to invoice, meeting request, newsletter, other, then branch on {{ $json.label }} with a Switch node — and send null results to a manual-review branch.

Troubleshooting

Section titled “Troubleshooting”
  • “Choose a zero-shot model to classify with.” — pick a model; install one from the picker or the Local AI page.
  • “Add at least one label (comma-separated).” — the Labels field is empty after parsing.
  • label is often null — lower the threshold, or make labels more distinct and descriptive (refund request rather than money).
  • Several labels look right — turn on Allow several labels and read the labels scores.
Section titled “Related nodes”