Typed-question APIs

Some models don’t chat. You send them some input and a list of typed questions, and they send back a typed answer to each one, such as true or false, one choice from a list, or a score. That’s useful for routing support tickets, flagging content, or scoring urgency without parsing free-form text. ngrok.ai serves two of these APIs:

API	Endpoint	Provider	Models
OpenAI Decisions	`POST /v1/decisions`	OpenAI	`gpt-6-luna`
Typesafe AI System One	`POST /v1/systemone`	Typesafe AI	`jev-latest`, `jev-preview`, `jev-1.13.0`

You can pay with credits for both, or bring your own provider key for OpenAI or Typesafe AI.

Decisions requests

A Decisions request has an input, in the same message format as the Responses API, and an array of questions. Each question has a type (predicate, choice, or score), a name, and instructions.

curl https://gateway.ngrok.ai/v1/decisions \
  -H "Authorization: Bearer ng-xxxxx-g1-xxxxx" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "gpt-6-luna",
    "input": [
      {
        "role": "user",
        "content": [
          {"type": "input_text", "text": "I was charged twice for my March invoice and need one charge refunded today."}
        ]
      }
    ],
    "questions": [
      {"type": "predicate", "name": "is_refund_request", "instructions": "Does the customer ask for a refund?"},
      {
        "type": "choice",
        "name": "team",
        "instructions": "Which team owns this ticket?",
        "choices": [
          {"value": "billing", "description": "Charges, invoices, and refunds."},
          {"value": "technical", "description": "Product defects and outages."}
        ]
      }
    ]
  }'

The response has an answers array with one entry per question. A predicate answer has the probability that it’s true. A choice answer names the choice it picked, with a probability for each option:

{
  "model": "gpt-6-luna",
  "answers": [
    {"type": "predicate", "name": "is_refund_request", "probability": 1.0},
    {
      "type": "choice",
      "name": "team",
      "choice": "billing",
      "probabilities": [
        {"value": "billing", "probability": 1.0},
        {"value": "technical", "probability": 0.0}
      ],
      "confidence": 1.0
    }
  ],
  "usage": {"input_tokens": 286, "output_tokens": 0, "total_tokens": 286}
}

The gateway forwards model, input, questions, and safety_identifier. It removes any other top-level field and names it in the Ngrok-AIG-Removed-Params response header. See OpenAI’s Decisions guide for every question type and option. "model": "ngrok/auto" also works and always picks gpt-6-luna. Any other model returns ERR_NGROK_27226, because no other model serves Decisions. When you pay with credits, OpenAI bills a decision on input tokens only, at gpt-6-luna’s input rate.

Typesafe AI requests

A System One request has a state, the text to evaluate, and a questions object that maps each question ID to a typed question.

curl https://gateway.ngrok.ai/v1/systemone \
  -H "Authorization: Bearer ng-xxxxx-g1-xxxxx" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "jev-latest",
    "state": "The server returned 500 for every request for ten minutes.",
    "questions": {
      "is_urgent": {
        "type": "noul",
        "instructions": "Is this an urgent production incident?"
      }
    }
  }'

The response maps each question ID to its answer, with a calibrated probability:

{
  "model": "jev-1.13.0",
  "answers": {"is_urgent": {"type": "noul", "noul": 0.91}},
  "usage": {"input_tokens": 287, "output_tokens": 23}
}

The gateway forwards model, state, and questions, and removes any other top-level field. See the Typesafe AI API reference for question types such as choice and score, and the Typesafe AI provider page for setup. Name a Jev model in the request. ngrok/auto doesn’t pick Jev models, so it returns ERR_NGROK_27227 on this endpoint.

What’s different from chat requests

Both APIs differ from Chat Completions and Messages in a few ways:

No streaming: A request with "stream": true returns ERR_NGROK_27228 before reaching the provider.
No fallback list: The gateway reads only model. Each API has a single provider, so a models list is removed rather than used.
POST only: Each API has a single create endpoint. Other methods and paths below it return an error.
Same billing model: Every request pays the processing fee. When you pay with credits, credits also cover the model cost.

Next steps

Typesafe AI: Set up Typesafe AI with credits or your own key
OpenAI: Set up OpenAI with credits or your own key
How it works: Read the gateway’s response headers

Concepts

Getting Started With

SDKs

Agents

Guides

Observability

Examples

Reference

API Reference

Decisions requests

Typesafe AI requests

What’s different from chat requests

Next steps

​Decisions requests

​Typesafe AI requests

​What’s different from chat requests

​Next steps

Decisions requests

Typesafe AI requests

What’s different from chat requests

Next steps