Classify steps

A classify step answers one question about its inputs with a classifier model (Jev). It returns probabilities, so a later script can apply a threshold. A classify step has exactly one of three forms. The shipped support-triage example uses all three:

  team:
    name: Choose the team
    in: {subject: ticket.subject, body: ticket.body}
    do:
      kind: classify
      question: Which team should handle the main request in this support ticket?
      options:
        billing: Billing - invoices, charges, refunds, and plan changes.
        technical: Technical - errors, outages, and problems using the product.
        account: Account - sign-in, users, and access.
        unclear: Unclear - the ticket does not say enough to choose, or no listed team fits.
    out: team

  refund:
    name: Check for a refund request
    in: {body: ticket.body}
    do:
      kind: classify
      question: Does the customer ask for money back?
      answer: yes_no
    out: refund

  urgency:
    name: Score the urgency
    in: {subject: ticket.subject, body: ticket.body}
    do:
      kind: classify
      question: How urgent is this support ticket for the customer?
      levels: [low, normal, high, urgent]
    out: urgency

Form	`do`	Output value
choice	`options: {ID: description}` (2 to 255 options)	`{choice: ID, probabilities: {ID: p}}`
yes/no	`answer: yes_no`	`{answer: boolean, probability: p}`, where p is the probability of yes
level	`levels: [NAME, ...]` (2 to 10 names, lowest first)	`{level: NAME, score: number, probabilities: {NAME: p}}`, where score is the expected level index (0 is the first level)

out is one name. Later steps read its fields, for example team.choice, team.probabilities.billing, or refund.probability.

Rules

Give the step a name, a question, and at least one input through in or each. The question and the option descriptions are literal text.
Inputs are JSON data. A file needs an earlier step that reads it.
A classify step cannot declare changes.
Add an option for unclear inputs, and decide in a script what to do with a low probability. In the example, route uses the team only when it is not unclear and its probability is at least 80%.
Counting, math, and date comparisons belong in a script, not in a question.

Where it runs

Classification goes through your Method account and uses the same model credit as hosted models. The run’s usage.cost_usd includes it. With method config model-key, classify steps on this computer call OpenRouter with your own key. Requests are private: no training and zero data retention. See Privacy. A request that the service briefly cannot answer gets three attempts.

​Rules

​Where it runs

Rules

Where it runs