DocsQuestions & answers

Choice, score and noul questions

A request carries one state (the text to judge) and up to 32 questions. Each question has one of three types. Every question is answered independently against the same state, in a single forward pass, so adding one more question never changes the others.

Explore the three types

Pick a type to see how Laya reaches its answer, what you send, and what comes back. Then run it on your own text.

choice: pick one label

Like a multiple-choice quiz. Laya scores every option against the text, the scores add up to 100%, and the most likely one becomes the answer.

Example answer

“What does the customer want?”

refund…

technical_help…

cancellation…

other…

→ choice = refund

Try it on your own text

Free demo, no key, processed in Switzerland.

You send (in questions)

choice questionjson

{
  "intent": {
    "type": "choice",
    "instructions": "What does the customer want?",
    "criteria": {
      "refund": "money back or a duplicate charge reversed",
      "technical_help": "a bug or integration problem",
      "cancellation": "wants to cancel",
      "other": "anything else"
    }
  }
}

You get back (answers.intent)

answer shapejson

{
  "type": "choice",
  "choice": "refund",                 // the most probable label
  "probabilities": { "refund": 0.9466, "technical_help": 0.0071, ... },
  "confidence": 0.7899,               // 1 − normalised entropy
  "action": { "act_probability": 1.0 }
}

Tips for choice

Describe each label in criteria (a short phrase each). Bare names give the model less to work with.
Always include an "other" option so unrelated text has somewhere to go.
Keep it small. Accuracy drops on big taxonomies (Banking77, 77 labels: 0.425), so split into a coarse and a fine question.

Reference: the three question types

choice — pick one label

choicejson

"intent": {
  "type": "choice",
  "instructions": "What does the customer want?",
  "criteria": {
    "refund": "money back or a duplicate charge reversed",
    "tracking": "where is my parcel",
    "other": "anything else"
  }
}
// criteria may also be a plain list: ["refund", "tracking", "other"]

score — ordinal levels

scorejson

"frustration": {
  "type": "score",
  "instructions": "How frustrated does the customer sound?",
  "criteria": ["calm", "concerned", "annoyed", "very angry"]
}
// answer.score is the expected level index (0-3 here), answer.probabilities the distribution

noul — is this statement true?

nouljson

"urgent": {
  "type": "noul",
  "instructions": "Does the message communicate time pressure?",
  "criteria": { "true": "a deadline or 'today/now' is stated", "false": "no time pressure" }
}
// criteria is optional; answer.noul is P(true)

Request fields

Field	Description
`state`string \| object \| array	What to decide about. Objects are serialised to JSON. Only the first 512 (English) or 1,024 (multilingual) tokens per question are read — put the decisive text first.
`questions`requiredobject	Map of your question ids to definitions. Max 32 per request, max 64 options per question.
`model`"english" \| "multilingual" \| "typed-decision…	Pin a checkpoint. Omitted (or an unknown id such as a Jev model name) → automatic routing.
`lang`string	Language hint ("en", "de", "pt-BR"). Beats automatic detection, which can mistake short non-English Latin-script text for English.

Choosing thresholds

Laya's checkpoints ship over-confident before per-task calibration (the model card reports mean ECE falling from 0.466 to 0.081 after temperature fitting). Treat confidence as a ranking signal: collect a few hundred labelled examples, find the confidence above which accuracy meets your bar, automate above it and send the rest to a human. See calibrated probabilities and act/escalate routing.

Tips for writing good questions

Describe each label in criteria; bare label names give the model less to work with.
Keep label sets small. Accuracy drops on large taxonomies (Banking77, 77 labels: 0.425) — split into a coarse question and a fine one.
Ask one thing per question. Each question reads the state once and is billed its input tokens, so ask only what you act on; five focused questions answer better than one compound one.
If a noul answer seems to follow its own true/false wording rather than the text, rephrase it as a two-option choice with neutral keys.
For the four typed-decision workflows (customer service, invoices, security incidents, agent traces), pass "model": "typed-decisions".