API¶

Endpoints¶

POST /v1/systemone¶

Request: {state, model, questions}.

• state: str | dict | list — plain text, a record, or a sequence of records.
• model: non-empty string, must be a known Jev name, an extra alias, or the serving model id. See Configuration.
• questions: non-empty map of caller-chosen ids to one question each.

Question types¶

choice — 1–255 options as a map (descriptions may be null):

{
  "type": "choice",
  "instructions": "Which team should handle this?",
  "criteria": {"billing": "Payments, invoicing, refunds", "technical": "Bugs, outages"}
}

score — 2–10 ordered levels as a list:

{
  "type": "score",
  "instructions": "How frustrated is the customer?",
  "criteria": ["Calm", "Frustrated", "Very angry"]
}

noul — yes/no/uncertain-likelihood in [0, 1]:

{
  "type": "noul",
  "instructions": "The message conveys urgency?",
  "criteria": {"true": "time-sensitive", "false": "no urgency"}
}

instructions and criteria descriptions accept str | dict | list (structured prompts); non-string values are rendered by the backend. noul criteria are optional.

Answers¶

Response: {model, answers, usage} with one typed answer per question id:

• choice: {type, choice, probabilities, confidence}.
• score: {type, score, legend, probabilities, confidence} with stringified level keys ({"0": …}) on the wire, matching Jev HTTP.
• noul: {type, noul} — no confidence field, matching Jev.

usage is {input_tokens, output_tokens}. output_tokens is always 0 (Laya never generates tokens); input_tokens comes from the backend.

GET /v1/models¶

Returns the serving checkpoint plus every accepted alias:

{"models": [{"name": "laya-english", "description": "...", "release_date": "unknown"}]}

Errors¶

All errors use {"error": {"message", "field"}}:

Validation limits: choice ≤ 255 options, score 2–10 levels, model required, non-empty questions. Context budgets mirror Jev's 64k (state + all questions) and 32k (state + longest question); oversize requests are 422. Sizes are counted with the checkpoint tokenizer when one is resident, word approximation otherwise.