OpenAI-compatible decisions

POST /v1/decisions answers the same questions as POST /v1/systemone, in the wire format of OpenAI's Decisions API. A client built for that API needs only a base URL and a model change: the same nace_sk_ key, funding, rate limits, validation and billing apply, and the request reaches the model as the same bytes a System One request would. Use it when your code already speaks OpenAI's Decisions shape. New integrations that want typed answers keyed by name can use the Nace SDKs and /v1/systemone directly.

Base URL and modelsCopy Anchor Link

Point the client at https://console.nace.ai/v1 and name a Drex model. drex-max and drex-max-v1.1 read text and images; drex-v2.0 and drex-latest read text. The Models reference lists them. model resolves as on /v1/systemone: a registered id or alias is served as named, no model is served like drex-latest, and any other name, gpt-6-luna and a mistyped drex- name included, returns 422 with code: model_not_found and param: model.

import OpenAI from "openai";

const client = new OpenAI({ baseURL: "https://console.nace.ai/v1", apiKey: process.env.NACE_API_KEY });

const { answers, usage } = await client.decisions.create({
  model: "drex-v2.0",
  input: "I was charged twice for my March invoice. Please refund one of the charges.",
  questions: [
    { type: "predicate", name: "wants_refund", instructions: "Is the customer asking for a refund?" },
    { type: "choice", name: "topic", instructions: "Which topic is it?", choices: [{ value: "billing" }, { value: "shipping" }, { value: "other" }] },
    { type: "score", name: "urgency", instructions: "How urgent is it?", levels: [{ label: "low" }, { label: "medium" }, { label: "high" }] },
  ],
});

from openai import OpenAI

client = OpenAI(base_url="https://console.nace.ai/v1", api_key=os.environ["NACE_API_KEY"])

decision = client.decisions.create(
    model="drex-v2.0",
    input="I was charged twice for my March invoice. Please refund one of the charges.",
    questions=[
        {"type": "predicate", "name": "wants_refund", "instructions": "Is the customer asking for a refund?"},
        {"type": "choice", "name": "topic", "instructions": "Which topic is it?", "choices": [{"value": "billing"}, {"value": "shipping"}, {"value": "other"}]},
    ],
)

LiteLLM reaches Drex through its OpenAI provider: a model entry openai/drex-max with api_base: https://console.nace.ai/v1 and your key as api_key. Vercel AI Gateway's /v1/decisions route takes the same base URL change.

How the request mapsCopy Anchor Link

Each OpenAI field becomes its System One counterpart before validation, so every rule on Questions, State and Drex Max applies unchanged.

ImagesCopy Anchor Link

Images are data URLs only (data:image/png, data:image/jpeg or data:image/webp, up to 16 MiB decoded and 32 MiB together); Drex does not fetch image URLs, and an https:// image_url returns 422 naming that part. Up to 8 per request, on drex-max and drex-max-v1.1 only: an image sent to drex-v2.0 returns 422 with code: invalid_value and param naming the first image part, the same refusal /v1/systemone gives as images: images require drex-max or drex-max-v1.1. Inline images count toward the 4 MiB body limit.

ResponseCopy Anchor Link

{
  "model": "drex-v2.0",
  "answers": [
    { "type": "predicate", "name": "wants_refund", "probability": 0.9868 },
    { "type": "choice", "name": "topic", "choice": "billing", "confidence": 0.986, "probabilities": [{ "value": "billing", "probability": 0.9895 }, { "value": "shipping", "probability": 0.0056 }, { "value": "other", "probability": 0.0049 }] },
    { "type": "score", "name": "urgency", "score": 1.53, "confidence": 0.72, "probabilities": [{ "value": 0, "label": "low", "probability": 0.16 }, { "value": 1, "label": "medium", "probability": 0.15 }, { "value": 2, "label": "high", "probability": 0.69 }] }
  ],
  "usage": {
    "input_tokens": 106,
    "output_tokens": 212,
    "total_tokens": 318,
    "input_tokens_details": { "cached_tokens": 0, "cache_write_tokens": 0 },
    "output_tokens_details": { "reasoning_tokens": 0 }
  },
  "request_id": "req_<32 hex characters>"
}

answers is in question order. usage.input_tokens is what Drex bills, image tokens included; the detail counts are always zero. request_id is Drex's; quote it in support requests. Drex's models answer every question, so no answer is a refusal today; the type is in OpenAI's contract and a client should keep handling it.

ErrorsCopy Anchor Link

Every error keeps Drex's error.type and HTTP status (see Errors and retries) and adds OpenAI's error.code and error.param. The OpenAI SDKs raise their usual classes: a 422 is UnprocessableEntityError with code, param and Drex's issues, each issue's path a param.

{
  "error": {
    "type": "invalid_request_error",
    "message": "questions[2].levels: must have 2 to 10 levels",
    "code": "invalid_value",
    "param": "questions[2].levels",
    "issues": [{ "path": "questions[2].levels", "message": "must have 2 to 10 levels" }]
  },
  "request_id": "req_<32 hex characters>"
}

Not supportedCopy Anchor Link

Streaming (OpenAI's Decisions API has none), per-question images or input_json inside instructions (a non-string instructions returns 422 naming questions[i].instructions), image URLs, more than 8 images, and messages from roles other than user.