Decide

Every field of POST /v1/decide: the request, the response, headers and body limits.

http

POST https://api.mellix.com/v1/decide

Answers one or more questions about one situation. It needs a bearer token and a JSON body with a Content-Length header (see Request bodies for compressed and multipart bodies). This endpoint serves the Myles-G models. The CPU models (Myles-400M, 150M and 22M) are not called over HTTP; see Model types. There is no text-generation endpoint.

Request body

Field	Type	Meaning
`state`	text or JSON	The situation to decide about. JSON is passed to the model as indented JSON. See State.
`questions`	object	One or more questions, each under a name you choose (up to 256, with up to 256 options each and 4,096 over all of them). Each answer comes back under the same name.
`assets`	object, optional	Images for the model to read (Myles-G 12B and Myles-G 31B), each under a reference you then write in the state, the instructions or the criteria descriptions. See Images.
`image_budget`	integer, optional	The most input tokens one image may use: `70`, `140` or `280` (the default). Lower is cheaper and faster but sees less detail. See Image budget.
`model`	text	A model id from `GET /v1/models`, for example `myles-g-12b` or `myles-g-31b`. Required: every request names its model. During a transition, a request without one is still answered by Myles-G 12B; this is being withdrawn, after which such a request is refused with `400` `model_required`. The response's `model` says which model answered.

Question object

Field	Type	Meaning
`type`	text	`noul`, `choice` or `score`.
`instructions`	text	The question, in plain language. `null` is read as no instructions.
`criteria`	object or list	For `choice`: labels mapped to descriptions, or a list of labels. For `score`: level descriptions, lowest first. Optional for `noul`: what `true` and `false` mean (or `yes` and `no`), both sides given. A description may be `null`: the option is then shown by its name (Yes or No, the label, or "Level 2"). Up to 256 options.

Response

Field	Meaning
`model`	The model that answered.
`answers`	One answer per question name. `noul`: the probability of yes; uncertain answers are committed, at least 0.80 towards the chosen side. `choice`: `choice`, `confidence` and `probabilities`. `score`: `level`, `score`, `confidence`, `legend` and `probabilities`. See Question types.
`truncated`, `state_chars`	Only when the model did not read the whole state: `truncated` is `true` and `state_chars` gives `original` (characters sent) and `kept` (characters the model read). A text state too long for the context window keeps its beginning and its end; a state with images keeps its first 20,000 characters. See State.
`usage`	`input_tokens` and `output_tokens` for the request. Images count as input tokens (see Billing). Hosted pricing counts input tokens only.

Example

bash

curl https://api.mellix.com/v1/decide \
  -H "Authorization: Bearer $MELLIX_API_TOKEN" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "myles-g-12b",
    "state": "Charged twice for order 4471, need the extra payment back",
    "questions": {
      "route": {"type": "choice", "instructions": "Which team?", "criteria": ["billing", "returns", "technical", "other"]}
    }
  }'

json

{
  "model": "myles-g-12b",
  "answers": {
    "route": {
      "type": "choice",
      "choice": "billing",
      "confidence": 0.89,
      "probabilities": { "billing": 0.93, "returns": 0.04, "technical": 0.01, "other": 0.02 }
    }
  },
  "usage": { "input_tokens": 118, "output_tokens": 1 }
}

The probabilities and token counts are illustrative. The response is one JSON object, sent when every question is answered; there is no streaming. To send an image with the request, see Images.

Response headers

Header	Meaning
`X-Request-Id`	An id for this call, on every response (errors included). Quote it to support.
`X-State-Truncated`	Only when the model did not read the whole state: `original=N; kept=M`, in characters, as in the response's `state_chars`.
`Mellix-Key-Id`, `Mellix-Key-Expires`	The calling key's id and expiry date, so a program can plan key rotation. See Authentication.
`Retry-After`	On `429` and `503`: the seconds to wait before trying again.

Request bodies

Size: up to 32 MiB as sent, and 32 MiB once decompressed.
Compression: send Content-Encoding: gzip or Content-Encoding: zstd with a compressed body. Any other encoding returns 415.
Multipart: a multipart/form-data body is accepted when its first part is named request and holds the JSON body. File parts carry images, named from assets; see Images.

Compatibility

The API works with existing decision-model clients:

POST /v1/systemone accepts the same request and returns the same response as POST /v1/decide. On that path a model the endpoint does not serve is, during the same transition, answered by Myles-G 12B; name a served model so the answer does not change when the transition ends.
Criteria given as an object with null descriptions, for example {"true": null, "false": null} or {"approve": null, "reject": null}, are read by option name.
choice and score answers carry confidence, and score answers carry legend, beside the fields described above. Clients that ignore fields they do not know see no change.
Compressed (gzip, zstd) and multipart/form-data bodies are accepted, as above.
GET /v1/models lists the models both as data and as models.
An image opt-in header that some clients send is accepted and not required.