TypeSafe → SIE

SIE serves TypeSafe’s System One API at POST /v1/systemone. Code written with the TypeSafe SDK runs on SIE unchanged: noul, choice and score questions, text, object and conversation states, models.list() and the async client. You point the SDK at SIE and name an SIE decision model.

client = TypeSafeClient(model="jev-latest")
client = TypeSafeClient(
    api_key="sk-sie-…",
    base_url="https://api.superlinked.com",
    model="fastino/GLiNER2.5-Decide",
)

export TYPESAFE_BASE_URL=https://api.superlinked.com
export TYPESAFE_API_KEY=$SIE_API_KEY
export TYPESAFE_DEFAULT_MODEL=fastino/GLiNER2.5-Decide

The Python and TypeScript SDKs both read the three TYPESAFE_* variables, so code that builds its client from the environment needs no edit at all. The base URL is the API root: the SDK adds /v1/systemone itself.

Before

Python
TypeScript

from typesafe_sdk import Choice, Noul, Score, TypeSafeClient

TICKET = "I was charged twice for my March invoice. Please refund one of the charges today."

with TypeSafeClient(model="jev-latest") as client:
    response = client.system_one(
        state=TICKET,
        questions={
            "spam": Noul(instructions="Is this message unsolicited spam?"),
            "intent": Choice(
                instructions="What does the customer want?",
                criteria={
                    "billing": "payments, invoices or refunds",
                    "technical": "something is broken",
                    "other": None,
                },
            ),
            "urgency": Score(
                instructions="How urgent is this request?",
                criteria=["can wait", "this week", "today"],
            ),
        },
    )

print(response.nouls["spam"].noul)
print(response.choices["intent"].choice)
print(response.scores["urgency"].score)

import { TypeSafeClient, choice, noul, score } from "@typesafe-ai/sdk";

const TICKET = "I was charged twice for my March invoice. Please refund one of the charges today.";

const client = new TypeSafeClient({ defaultModel: "jev-latest" });
const { answers } = await client.systemOne({
  state: TICKET,
  questions: {
    spam: noul("Is this message unsolicited spam?"),
    intent: choice("What does the customer want?", {
      billing: "payments, invoices or refunds",
      technical: "something is broken",
      other: null,
    }),
    urgency: score("How urgent is this request?", ["can wait", "this week", "today"]),
  },
});

console.log(answers.spam.noul, answers.intent.choice, answers.urgency.score);

After

Only the client changes. The questions and the way you read the answers stay as they are.

Python
TypeScript

from typesafe_sdk import Choice, Noul, Score, TypeSafeClient

TICKET = "I was charged twice for my March invoice. Please refund one of the charges today."

with TypeSafeClient(
    api_key="sk-sie-…",
    base_url="https://api.superlinked.com",
    model="fastino/GLiNER2.5-Decide",
) as client:
    response = client.system_one(
        state=TICKET,
        questions={
            "spam": Noul(instructions="Is this message unsolicited spam?"),
            "intent": Choice(
                instructions="What does the customer want?",
                criteria={
                    "billing": "payments, invoices or refunds",
                    "technical": "something is broken",
                    "other": None,
                },
            ),
            "urgency": Score(
                instructions="How urgent is this request?",
                criteria=["can wait", "this week", "today"],
            ),
        },
    )

print(response.nouls["spam"].noul)
print(response.choices["intent"].choice)
print(response.scores["urgency"].score)

import { TypeSafeClient, choice, noul, score } from "@typesafe-ai/sdk";

const TICKET = "I was charged twice for my March invoice. Please refund one of the charges today.";

const client = new TypeSafeClient({
  apiKey: "sk-sie-…",
  baseURL: "https://api.superlinked.com",
  defaultModel: "fastino/GLiNER2.5-Decide",
});
const { answers } = await client.systemOne({
  state: TICKET,
  questions: {
    spam: noul("Is this message unsolicited spam?"),
    intent: choice("What does the customer want?", {
      billing: "payments, invoices or refunds",
      technical: "something is broken",
      other: null,
    }),
    urgency: score("How urgent is this request?", ["can wait", "this week", "today"]),
  },
});

console.log(answers.spam.noul, answers.intent.choice, answers.urgency.score);

Choosing a model

/v1/systemone 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 state
`fastino/GLiNER2.5-Decide-1B`	2,048 tokens	All of the questions (at most 512 tokens) and the state; 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 state, 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.

What differs from TypeSafe’s hosted API

Model names. SIE takes its own model ids and has no aliases such as jev-latest. Set the model on the client, in TYPESAFE_DEFAULT_MODEL, or per call with model=.
Score levels. A score question takes 1 to 10 levels.
Input that does not fit. SIE refuses a state or question set that does not fit the model’s window with 422, which the Python SDK raises as TypeSafeUnprocessableEntityError. It never truncates text. Use a model with a longer window for long states.
Other models. models.list() lists every model the server serves. A system_one call with a model that does not answer typed questions gets 422.
Loading. A model that is still loading answers 503 with a Retry-After header.

Run it yourself

GLiNER2.5-Decide is a 340M-parameter encoder, so a CPU serves it:

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

export TYPESAFE_BASE_URL=http://localhost:8080
export TYPESAFE_DEFAULT_MODEL=fastino/GLiNER2.5-Decide
export TYPESAFE_API_KEY=local   # a local server checks no key

The first request downloads and loads the model. The systemone-local example in the SIE repository runs the TypeSafe SDK through every feature that reaches the server and reports each case.

Also calling OpenAI’s Decisions API? See OpenAI Decisions → SIE.