OpenAI Decisions → SIE

SIE serves OpenAI’s Decisions format at POST /v1/decisions. The OpenAI SDK’s client.decisions.create runs on SIE with two changes: the client’s base URL and key, and the model. Predicate, choice and score questions keep their shape, and the answers come back in question order, as they do from OpenAI.

client = OpenAI()
client = OpenAI(api_key="sk-sie-…", base_url="https://api.superlinked.com/v1")

const client = new OpenAI();
const client = new OpenAI({ apiKey: "sk-sie-…", baseURL: "https://api.superlinked.com/v1" });

Then model="gpt-6-luna" becomes model="fastino/GLiNER2.5-Decide", or another decision model.

Before

Python
TypeScript

from openai import OpenAI

client = OpenAI()
decision = client.decisions.create(
    model="gpt-6-luna",
    input="I was charged twice. Please refund one charge today.",
    questions=[
        {"type": "predicate", "name": "refund", "instructions": "Does the customer request a refund?"},
        {
            "type": "choice",
            "name": "team",
            "instructions": "Which team should handle this?",
            "choices": [{"value": "billing"}, {"value": "technical"}],
        },
        {
            "type": "score",
            "name": "urgency",
            "instructions": "How urgent is the request?",
            "levels": [{"label": "can wait"}, {"label": "today"}],
        },
    ],
)
refund, team, urgency = decision.answers
print(refund.probability, team.choice, urgency.score)

import OpenAI from "openai";

const client = new OpenAI();
const decision = await client.decisions.create({
  model: "gpt-6-luna",
  input: "I was charged twice. Please refund one charge today.",
  questions: [
    { type: "predicate", name: "refund", instructions: "Does the customer request a refund?" },
    {
      type: "choice",
      name: "team",
      instructions: "Which team should handle this?",
      choices: [{ value: "billing" }, { value: "technical" }],
    },
    {
      type: "score",
      name: "urgency",
      instructions: "How urgent is the request?",
      levels: [{ label: "can wait" }, { label: "today" }],
    },
  ],
});
console.log(decision.answers);

After

Python
TypeScript

from openai import OpenAI

client = OpenAI(api_key="sk-sie-…", base_url="https://api.superlinked.com/v1")
decision = client.decisions.create(
    model="fastino/GLiNER2.5-Decide",
    input="I was charged twice. Please refund one charge today.",
    questions=[
        {"type": "predicate", "name": "refund", "instructions": "Does the customer request a refund?"},
        {
            "type": "choice",
            "name": "team",
            "instructions": "Which team should handle this?",
            "choices": [{"value": "billing"}, {"value": "technical"}],
        },
        {
            "type": "score",
            "name": "urgency",
            "instructions": "How urgent is the request?",
            "levels": [{"label": "can wait"}, {"label": "today"}],
        },
    ],
)
refund, team, urgency = decision.answers
print(refund.probability, team.choice, urgency.score)

import OpenAI from "openai";

const client = new OpenAI({ apiKey: "sk-sie-…", baseURL: "https://api.superlinked.com/v1" });
const decision = await client.decisions.create({
  model: "fastino/GLiNER2.5-Decide",
  input: "I was charged twice. Please refund one charge today.",
  questions: [
    { type: "predicate", name: "refund", instructions: "Does the customer request a refund?" },
    {
      type: "choice",
      name: "team",
      instructions: "Which team should handle this?",
      choices: [{ value: "billing" }, { value: "technical" }],
    },
    {
      type: "score",
      name: "urgency",
      instructions: "How urgent is the request?",
      levels: [{ label: "can wait" }, { label: "today" }],
    },
  ],
});
console.log(decision.answers);

What carries over

Input. A string, or text messages with user, developer, system or assistant roles.
Questions. Predicate, choice and score questions, up to 200 per request within the model’s window. Answers keep the order of the questions, including unnamed questions and repeated names.
Choices. Up to 255 per question. A value is a string or a boolean, and true stays distinct from "true".
Scores. 1 to 10 levels. An answer carries each level’s probability and the expected level.
Errors. OpenAI error envelopes and retry headers, so the SDK raises its usual exception classes.

What differs

Models. SIE model ids replace OpenAI’s. The probabilities come from the SIE model you choose.
Text only. SIE’s decision models read text. A request with images, audio or tool calls gets a 400.
Input that does not fit. A request longer than the model’s window gets a 400. SIE never truncates it. Use a model with a longer window.

Choosing a model

/v1/decisions answers with SIE’s decision models. Each reads a fixed window of tokens per row:

Model	Window	What shares it
`fastino/GLiNER2.5-Decide`	512 tokens	All of a request’s questions (at most 256 tokens) and the input
`fastino/GLiNER2.5-Decide-1B`	2,048 tokens	All of the questions (at most 512 tokens) and the input; for longer text
`fastino/GLiNER2.5-multi-Decide`	2,048 tokens	As Decide-1B, multilingual
`convaiinnovations/laya-typed-decisions`	1,024 tokens	One question and the input, per question
`convaiinnovations/laya-multilingual`	1,024 tokens	As above, multilingual
`convaiinnovations/laya`	512 tokens	As above

GLiNER2.5-Decide runs on SIE Serverless. Self-hosted SIE serves all six.

Run it yourself

docker run -p 8080:8080 ghcr.io/superlinked/sie-server:latest-cpu-transformers5

Point the client at http://localhost:8080/v1 with any key: a local server checks none. The first request downloads and loads the model. The decisions-local example in the SIE repository checks every answer type, question ordering, the async client and the SDK’s error classes.

Using TypeSafe’s System One API instead? See TypeSafe → SIE. For OpenAI embeddings, see OpenAI → SIE.