Decisions

Answer typed questions about an input with any model of a project, through an endpoint compatible with OpenAI's Decisions API.

POST /v1/decisions answers typed questions about an input, with probabilities. It is compatible with OpenAI's Decisions API.

Any model in the Spicepod of the project can answer:

Model	Example `from`	Probabilities
TypeSafe decision model	`typesafe:jev`	Calibrated by the model.
OpenAI decision model	`openai:gpt-6-luna`	The model's estimates.
Chat model	`openai:gpt-4o-mini`, `anthropic:claude-haiku-4-5`	The model's own estimates, which are not calibrated.

Create a decision

POST https://data.spiceai.io/v1/decisions

Headers

Header	Type	Description
`X-API-Key`	string	Your project API key
`Content-Type`	string	`application/json`

Body

Field	Type	Description
`model`	string	The name of a model under `models` in the Spicepod. Required.
`input`	string or array	The text that every question is about, or user messages with text parts. Image parts return `400`. Required.
`questions`	array	1 to 200 questions, answered in order. Required.
`reasoning_effort`	string	`none`, `minimal`, `low`, `medium`, `high`, or `xhigh`. Spice passes it to a chat model. A decision model returns `400` for it. A Spice extension.
`safety_identifier`	string	An opaque ID for the end user, at most 128 characters. Spice forwards it to OpenAI decision models.

A field that this table does not list returns 400.

input can also be a list of user messages. Each message has the user role, and its content is a string or a list of input_text parts. An input_image part returns 400.

"input": [
  {"role": "user", "content": [{"type": "input_text", "text": "Help! My payouts have been failing for 3 days."}]}
]

Each question has a type, its instructions, and an optional name that its answer repeats:

`type`	Asks	Adds
`predicate`	How likely a statement about the input is true.	Nothing.
`choice`	Which option fits the input.	`choices`: 2 to 255 distinct options, each a `value` with an optional `description`.
`score`	Where the input sits on ordered levels.	`levels`: 2 to 10 levels, lowest first, each a `label` with an optional `description`.

curl -X POST "https://data.spiceai.io/v1/decisions" \
  -H "X-API-Key: <project-api-key>" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "jev",
    "input": "Help! My payouts have been failing for 3 days.",
    "questions": [
      {"type": "predicate", "name": "urgent", "instructions": "Does this convey urgency?"},
      {"type": "choice", "name": "team", "instructions": "Which team should handle this?",
       "choices": [{"value": "billing", "description": "Payments and payouts"}, {"value": "technical", "description": "Bugs and outages"}]},
      {"type": "score", "name": "tone", "instructions": "How upset is the customer?",
       "levels": [{"label": "calm"}, {"label": "annoyed"}, {"label": "furious"}]}
    ]
  }'

Response

answers holds one answer per question, in question order. Each answer has the question's type and name, or null when the question has no name.

`type`	Fields
`predicate`	`probability`: the probability that the statement is true.
`choice`	`choice`: the most probable option. `probabilities`: each option's probability. `confidence`: 0 for a uniform distribution, 1 for a certain one.
`score`	`score`: the probability-weighted, 0-based index of the levels. `probabilities` and `confidence`, as for `choice`.
`refusal`	None. The model declined this question, and the other answers stand.

{
  "model": "jev",
  "answers": [
    {"type": "predicate", "name": "urgent", "probability": 0.9},
    {"type": "choice", "name": "team", "choice": "billing",
     "probabilities": [{"value": "billing", "probability": 0.8}, {"value": "technical", "probability": 0.2}], "confidence": 0.6},
    {"type": "score", "name": "tone", "score": 1.3,
     "probabilities": [{"value": 0, "label": "calm", "probability": 0.1}, {"value": 1, "label": "annoyed", "probability": 0.5}, {"value": 2, "label": "furious", "probability": 0.4}], "confidence": 0.3}
  ],
  "usage": {"input_tokens": 612, "input_tokens_details": {"cached_tokens": 0, "cache_write_tokens": 0},
            "output_tokens": 0, "output_tokens_details": {"reasoning_tokens": 0}, "total_tokens": 612}
}

model is the name that the request used. usage is absent when the model reports no token counts.

Errors

Errors use OpenAI's envelope: {"error": {"message", "type", "param", "code"}}.

Status	Cause
`400`	The request is not valid. Unknown fields, image inputs, and `reasoning_effort` for a decision model return it.
`401`	The model provider rejected the credentials.
`403`	The model provider denied access.
`404`	No model has this name.
`429`	The request was rate limited.
`500`	The model could not answer.
`503`	The model provider is unavailable.

In the Portal

The Playground has a Decisions page that sends these requests to the project. See Playground.