systemone API¶

POST /v1/systemone takes one request, a state and the questions about it, and returns one typed answer per question.

Request¶

Field	Type	What it holds
`state`	any JSON	what the questions are about: a string, an object or a list
`questions`	object	one question per key; the key names its answer
`model`	string	optional; the served model's name or one of its aliases
`media`	list	OneJev only: images and videos the state refers to by `<image:N>` and `<video:N>`
`images`	list	chat models: data URLs or base64, shown ahead of the text

Extra keys are accepted and ignored. A request needs at least one question.

{
  "state": "Order #1182 arrived with a cracked screen. The customer asks for a refund and attached two photos.",
  "questions": {
    "action": {"type": "choice", "instructions": "What should support do next?",
               "criteria": {"refund": "refund the order", "replace": "send a replacement", "escalate": "ask a human"}},
    "urgent": {"type": "noul", "instructions": "Does this need a reply within the hour?"}
  }
}

Questions¶

Every question has a type, instructions and, for a choice or a score, criteria.

Type	`criteria`	Answer
`noul`	optional `{"true": ..., "false": ...}`, what each outcome means	the probability that the statement holds
`choice`	object of option name to description, up to 256 options	the most likely option and every option's probability
`score`	list of levels, up to 256	the expected level and every level's probability

Instructions, descriptions and levels may be strings or any JSON. Each model family reads up to its own number of options and levels, given on its page in Supported models; a larger question gets a 422.

"severity": {"type": "score", "instructions": "How serious is the problem?",
             "criteria": ["cosmetic", "annoying", "product unusable"]}

Response¶

Qwen3.5-4B answered the request above with these answers. The server's own record under meta is left out.

{
  "model": "Qwen3.5-4B",
  "answers": {
    "action": {
      "type": "choice",
      "choice": "refund",
      "probabilities": {
        "refund": 0.777006,
        "replace": 0.208265,
        "escalate": 0.014729
      },
      "confidence": 0.665509
    },
    "urgent": {
      "type": "noul",
      "noul": 0.827155
    }
  },
  "usage": {
    "input_tokens": 110,
    "output_tokens": 2
  }
}

Answer field	Type	Meaning
`noul`	number	probability that the statement holds
`choice`	string	the option with the highest probability
`score`	number	expected level, from 0 to the number of levels minus 1
`probabilities`	object	each option's, or each level's, probability; they sum to 1
`confidence`	number	0 when the probabilities are uniform, 1 when one option holds them all
`legend`	object	the levels of a score by their index

usage counts the tokens the request took, the state once plus every question, as input_tokens, and one output token per question. The response also carries the server's own record under meta (onejev for OneJev): the time spent planning, waiting and computing, and whether the request ran in a CUDA graph.

A choice's confidence is its top probability rescaled from uniform to certain, (p_max − 1/k) / (1 − 1/k) for k options. A score's confidence is 1 minus the expected distance from the most likely level, divided by the same distance for a uniform spread.

Query parameters¶

Parameter	What it does
`debias=N`	OneJev: asks each question in N cyclic shifts of its options and combines them

Errors¶

Status	When
422	the body is not valid JSON, a field has the wrong shape, a question has more options than the model reads, or the request is too long for the model
404	`model` names a model this server does not serve
500	the server failed while answering

A 422 lists the problems in detail, each with the place in the body (loc) and a message (msg). A request with images on a machine without torchvision gets a 500 whose detail says how to install it.

Other endpoints¶

Endpoint	What it returns
`GET /health`	`{"status": "ok"}` with the model, its family and the Flide version; 503 while the model loads
`GET /v1/models`	the served model and its aliases
`POST /flush_cache`	empties the session cache (`GET` works too)
`GET /`	a page to try requests in the browser