Custom Classifiers

A classifier is a lightweight rule that categorizes text by keyword. You define categories once in your account and use them to tag and triage content.

Available now (Beta) in the dashboard. Go to Settings → Classifiers (/dashboard/settings/classifiers) to create categories (a name plus match terms) and classify any text against them live — the tester returns every category whose terms appear. Classification is deterministic (keyword match, no model call), so it's fast and predictable.

What classifiers do today#

• Categorize on demand — paste or send any text to the classify tester and get back every matching category.
• Deterministic matching — categories fire on their match terms, with no model call, no latency, and no inference cost.
• Account-scoped — categories are defined once and shared across your workspace.

Use cases#

Defining a classifier#

Create categories at Settings → Classifiers: give each category a name and one or more match terms, then use the built-in tester to classify sample text and confirm the categories fire as expected.

Attaching classifiers to requests#

Not yet available. Auto-attaching classifiers to every request as a pipeline step — labels on logs, block/retry gating, and a model-graded classification mode (target, on_match actions, response labels under x_opendunes) — is tracked to be built. Today you manage categories and classify on demand from the dashboard; a classifiers field in a chat request is ignored.

Best practices#

• Keep match terms specific. Short generic words over-fire; multi-word terms and distinctive vocabulary keep categories precise.
• Test with real samples. Run representative production text through the tester before relying on a category for triage.
• Combine with logging. Pair classification with the activity log when reviewing how flagged conversations actually went.