Usage

d1-omni-600M · ONNX on CPU · no auth

d1-omni is a decision model: it does not write text. You give it a state (the thing being decided about), a typed question and, optionally, an image; it returns a probability distribution over the question's options from a single forward pass.

Endpoints

Method & path	What it does
`POST /api/decide`	The inference route — everything below is about this one.
`GET /api/health`	200 + `{"status": "ok"}` once the model is loaded.
`GET /api/info`	Model files, quant levels and limits.
`GET /docs`	Interactive OpenAPI docs (try it in the browser).

Request

JSON body, content-type: application/json:

Field	Type	Notes
`question`	object	Required. `type`, `instructions`, `criteria` — see below.
`state`	string or any JSON	What the question is about. Optional with an image; defaults to empty text.
`image`	string	Optional. Base64 of the image bytes, or a full `data:image/png;base64,...` URL.

Question types

`type`	`criteria`	Returns
`choice`	`{name: description}`, at least 2 options	`choice` (winning option name), `confidence`, `probabilities` per option
`noul`	optional `{"true": "...", "false": "..."}` (or yes/no keys)	`noul` = P(yes), plus `probabilities`
`score`	list of 2–10 level descriptions, lowest first	`score` = expected level (0-based float), `probabilities` per level, `legend`

The descriptions matter: the model reads name: description for each option, so put the meaning in the description and keep names short.

Examples

Text · choice

curl -s https://d1.corbia.net/api/decide \
  -H 'content-type: application/json' \
  -d '{
    "state": "I was charged twice this month, please refund one of them.",
    "question": {
      "type": "choice",
      "instructions": "Which team should handle this?",
      "criteria": {
        "billing": "Charges, refunds, invoices",
        "technical": "App or site faults",
        "fraud": "Suspected unauthorised use"
      }
    }
  }'

Text · noul (yes/no)

curl -s https://d1.corbia.net/api/decide \
  -H 'content-type: application/json' \
  -d '{
    "state": {"subject": "Quarterly report", "sender": "unknown@extarnal-partner.com", "attachments": ["report.xlsm"]},
    "question": {"type": "noul", "instructions": "Is this email suspicious?"}
  }'

Text · score

curl -s https://d1.corbia.net/api/decide \
  -H 'content-type: application/json' \
  -d '{
    "state": "The API works but the docs are thin and two endpoints 500 under load.",
    "question": {
      "type": "score",
      "instructions": "Rate this developer review",
      "criteria": ["hostile", "unhelpful", "mixed", "helpful", "excellent"]
    }
  }'

Image · base64

# macOS:  base64 -i photo.jpg | pbcopy      Linux:  base64 -w0 photo.jpg > photo.b64
curl -s https://d1.corbia.net/api/decide \
  -H 'content-type: application/json' \
  -d "{
    \"question\": {
      \"type\": \"choice\",
      \"instructions\": \"What is in the picture?\",
      \"criteria\": {\"sign\": \"A sign\", \"receipt\": \"A receipt\", \"person\": \"A person\", \"food\": \"Food\"}
    },
    \"image\": \"$(cat photo.b64)\"
  }"

With an image, state is optional extra context (e.g. a caption or surrounding conversation). noul questions over images answer against the default false: no / true: yes wording. Phrase them as plain statements ("Is there a square?", "Is this a cat?") — yes/no answers get brittle with unusual attribute combinations (e.g. exact colours); prefer a choice when in doubt.

Python

import base64, requests

r = requests.post("https://d1.corbia.net/api/decide", json={
    "state": "Where is the traveller flying to?",
    "question": {
        "type": "choice",
        "instructions": "Which city?",
        "criteria": {"LAX": "Los Angeles", "SFO": "San Francisco", "JFK": "New York"},
    },
    "image": base64.b64encode(open("boarding-pass.png", "rb").read()).decode(),
})
print(r.json())

Response

{
  "type": "choice",
  "choice": "billing",
  "confidence": 0.81,
  "probabilities": {"billing": 0.81, "technical": 0.13, "fraud": 0.06},
  "meta": {"model": "d1-omni-600M", "modality": "text", "latency_ms": 180.4}
}

Text questions are temperature-calibrated (calmer, sharper distributions); image questions are not, matching the upstream runtime. One request = one forward pass; state is truncated to fit the context (16384 tokens for text, 896 alongside an image). Errors come back as 400 with a detail message.