TypeSafe Decisions

Decisions uses Jev to evaluate text or business state and returns answers that you can read directly in code. The API returns JSON synchronously and does not support streaming.

POST/v1/decisions

Authorization

bearerAuth

AuthorizationBearer <token>

Use an API Key issued by the current site.

In: header

Request Body

application/json

TypeScript Definitions

Use the request body type in TypeScript.

model*string

ID of a decision model available on the current site.

state*string||array<unknown>

Content or business state to evaluate. All questions in a request share the same state.

questions*

A nonempty question map. Each key is a custom question ID; answers use the same IDs.

Properties1 <= properties

stream?false

Omit or set to false. This API does not support stream: true.

Value in

false

Response Body

`application/json`

curl -X POST "https://example.com/v1/decisions" \  -H "Authorization: Bearer your-api-key" \  -H "Content-Type: application/json" \  -d '{    "model": "jev-1.13.0",    "state": "My subscription was charged twice. Please refund the duplicate charge.",    "questions": {      "ticket": {        "type": "noul",        "instructions": "Does the customer explicitly request a refund?",        "criteria": {          "true": "The message explicitly asks for a refund.",          "false": "The message does not ask for a refund."        }      }    }  }'

{
  "model": "jev-1.13.0",
  "answers": {
    "ticket": {
      "type": "noul",
      "noul": 0.99
    }
  },
  "usage": {
    "input_tokens": 319,
    "output_tokens": 20
  }
}

Request examples

The following examples use jev-1.13.0. Select a question scenario and code language to copy the corresponding request.

Determine whether a refund is requested

curl --request POST 'https://api.agtcloud.cn/v1/decisions' \
  --header 'Authorization: Bearer your-api-key' \
  --header 'Content-Type: application/json' \
  --data '{
  "model": "jev-1.13.0",
  "state": "My subscription was charged twice. Please refund the duplicate charge.",
  "questions": {
    "ticket": {
      "type": "noul",
      "instructions": "Does the customer explicitly request a refund?",
      "criteria": {
        "true": "The message explicitly asks for a refund.",
        "false": "The message does not ask for a refund."
      }
    }
  }
}'

Question types

Type	Purpose	Returned fields
`noul`	Determine whether a condition holds, such as whether a customer requests a refund	`noul`: probability that the condition holds
`choice`	Select a category or action from candidates, such as the team responsible for a ticket	`choice`, `probabilities`, `confidence`
`score`	Score against ordered criteria, such as ticket urgency	`score`, `legend`, `probabilities`, `confidence`

Provide the content to evaluate in state and the questions and criteria in questions. A request can contain multiple questions and mix the three types; every question shares the same state.

Reading results

answers uses the same question IDs as questions. For example, if the question ID is ticket, read the answer from answers.ticket.

Judgment: Read noul. Values range from 0–1; near 1 indicates a tendency toward true, and near 0 toward false. Set a decision threshold according to your business needs.
Classification: Read choice for the selected option name, and use probabilities to inspect each option's probability.
Scoring: Read score. Levels are numbered from 0 by their array position in criteria; three levels correspond to 0–2, and scores can be fractional. legend maps level numbers to their descriptions.

For choice and score, confidence indicates the certainty of the probability distribution, not the accuracy of the answer. Probabilities, scores, and Token usage vary with each request.

Tool selection uses choice to return an action name, such as lookup_payment. Your application performs the subsequent operation based on that name; the decision request itself does not execute tools or generate OpenAI tool_calls.

Token usage and costs

usage.input_tokens and usage.output_tokens record input and output Token usage respectively. Jev charges for input Tokens only; output Tokens are not charged. Multiple questions and candidate descriptions also affect input usage. For current prices, see the Model Marketplace.

200application/json

400application/json

422application/json

`application/json`

`application/json`

`application/json`