Decisions

Endpoint

POST https://jevmodel.io/api/v1/decisions
Content-Type: application/json
Authorization: Bearer sk-your-key-here

The endpoint accepts one task at a time and returns one decision.

Request body

Field	Type	Required	Description
`input`	string	yes	The text the decision is about. `state` is accepted as an alias for the same field.
`questions`	array or object	yes	At least one question. See question types.
`task`	string	no	A label for this call, used in your request history and in the credit ledger. Defaults to `support_triage`.
`schema`	any	no	Accepted for forward compatibility. It is not applied to the decision today.

{
  "task": "lead_scoring",
  "input": "We run a 40-person agency looking for a CRM with a REST API.",
  "questions": [
    {
      "key": "fit",
      "type": "score",
      "prompt": "How strong is the account fit?",
      "criteria": ["1", "2", "3", "4", "5"]
    },
    {
      "key": "route",
      "type": "choice",
      "prompt": "Should this lead go to sales or nurture?",
      "criteria": {
        "sales": "Ready to talk to a salesperson",
        "nurture": "Interested, but not ready yet"
      }
    }
  ]
}

questions accepts two shapes, and both produce the same result:

{ "questions": [ { "key": "fit", "type": "score", "prompt": "..." } ] }

{ "questions": { "fit": { "type": "score", "prompt": "..." } } }

In the object form the property name is used as the question key.

Question types

Every question has a key (yours, echoed back in the response), a prompt that describes the judgement to make, and a type.

`choice`

The model picks exactly one of your labels. criteria is required and must be an object that maps each label to a short description of what it means.

{
  "key": "route",
  "type": "choice",
  "prompt": "Should this lead go to sales or nurture?",
  "criteria": {
    "sales": "Ready to talk to a salesperson",
    "nurture": "Interested, but not ready yet"
  }
}

The answer carries the chosen label in choice, the per-label probabilities, and a confidence.

`score`

The model returns a value from the range you define. criteria is required and must be an array — the values it may answer with.

{
  "key": "fit",
  "type": "score",
  "prompt": "How strong is the account fit?",
  "criteria": ["1", "2", "3", "4", "5"]
}

The answer carries the numeric score, the legend that maps the numbers back to your values, the probabilities, and a confidence.

`noul`

The model estimates the probability that the input cannot be answered at all. noul questions take no criteria — any that you send are ignored.

{
  "key": "answerable",
  "type": "noul",
  "prompt": "Can this be answered from the text alone?"
}

The answer is a single number: the no-answer probability. It is also surfaced at the top level of the result as no_answer_probability.

A noul question never becomes the headline decision

The top-level decision prefers a labelled answer. If your request contains only a noul question, decision is the literal string noul and no confidence is returned, because the model produced none.

Response

{
  "code": 0,
  "message": "ok",
  "data": {
    "id": "jev_1761234567890",
    "object": "jev.decision",
    "created": 1761234567,
    "model": "jev-1.14",
    "provider": "TypeSafe",
    "upstream_id": "gen-1761234567-abc",
    "upstream_model": "typesafe/jev-1.14-20261014",
    "result": {
      "decision": "sales",
      "confidence": 0.88,
      "reason": "Probabilities: {\"sales\":0.88,\"nurture\":0.12}",
      "latency_ms": 1740,
      "no_answer_probability": 0.04,
      "answers": {
        "fit": {
          "type": "score",
          "score": 4,
          "legend": { "1": "1", "2": "2", "3": "3", "4": "4", "5": "5" },
          "probabilities": { "1": 0.01, "2": 0.03, "3": 0.08, "4": 0.88, "5": 0 },
          "confidence": 0.88
        },
        "route": {
          "type": "choice",
          "choice": "sales",
          "probabilities": { "sales": 0.88, "nurture": 0.12 },
          "confidence": 0.88
        }
      },
      "source": "live",
      "upstream_usage": {
        "cost": 0.0004,
        "input_tokens": 812,
        "output_tokens": 96
      }
    },
    "usage": {
      "credits": 1,
      "remaining_credits": 999,
      "daily_remaining": 99,
      "cost": 0.0004,
      "input_tokens": 812,
      "output_tokens": 96
    }
  }
}

Top level

Field	Description
`id`	Identifier for this decision, in the form `jev_<timestamp>`.
`object`	Always `jev.decision`.
`created`	Unix timestamp in seconds.
`model`	The Jev model that produced the decision, e.g. `jev-1.14`.
`provider`	The upstream provider, e.g. `TypeSafe`.
`upstream_id`	The upstream request id, for support and traceability.
`upstream_model`	The exact upstream model build.

`result`

Field	Description
`decision`	The primary decision: the chosen label, the score, or `noul`.
`confidence`	Confidence reported by the model. Omitted when the model reports none.
`reason`	A short human-readable explanation of the decision.
`latency_ms`	How long the decision took, in milliseconds.
`no_answer_probability`	The no-answer probability, when the request contained a `noul` question.
`answers`	The full answer set, keyed by your question keys.
`source`	Always `live` for calls served by the API.
`upstream_usage`	Cost and token accounting reported by the upstream.

`usage`

Field	Description
`credits`	Credits this call consumed. Always `1`.
`remaining_credits`	Credits left on the account after the call.
`daily_remaining`	Decisions left in today's allowance. Only present while you are on the gifted credits.
`cost`	Upstream cost in USD, when the provider reports it.
`input_tokens` / `output_tokens`	Upstream token accounting.

Validation rules

The API validates the request before it spends anything, and the messages are meant to be read literally:

questions must contain at least one question.
Every question needs a prompt (or instructions).
A choice question needs a non-empty criteria object.
A score question needs a non-empty criteria array.
A noul question ignores criteria — anything you send there is dropped before the call is made.
type must be one of choice, score, noul.

A request that breaks one of these rules returns 400 and is not billed.

On this page