Requests and responses

The CLI accepts UTF-8 JSON Lines: every non-empty line is one complete request. Jet returns one line for each request, preserves input order, and exits with a nonzero status if any line fails.

Request object

Each request contains two fields:

Field	Type	Meaning
state	any JSON value	Context shared by every question in the request.
questions	object	Named questions. Each key is preserved in answers.

The request deliberately has no model field. The CLI loads one model for the input stream using --model-path and identifies it with --model-id.

Question types

noul

A noul question returns a boolean-style decision weight. Criteria are optional; when present, they give the two values semantic meaning.

{
  "type": "noul",
  "instructions": "Should this incident page the on-call engineer?",
  "criteria": {
    "false": "Do not page",
    "true": "Page now"
  }
}

The response contains noul, the normalized weight assigned to the true candidate:

{"type":"noul","noul":0.91}

choice

A choice question maps stable keys to the semantic candidates shown to the model:

{
  "type": "choice",
  "instructions": "Choose the team that should own this ticket",
  "criteria": {
    "app": "Application team",
    "infra": "Infrastructure team"
  }
}

The response includes the winning key and every normalized candidate weight:

{"type":"choice","choice":"infra","probabilities":{"app":0.08,"infra":0.92}}

score

A score question receives an ordered list. The result is the probability-weighted, zero-based index across that list.

{
  "type": "score",
  "instructions": "Rate incident severity",
  "criteria": ["low", "medium", "high"]
}

{"type":"score","score":1.78,"probabilities":{"0":0.03,"1":0.16,"2":0.81}}

Successful response

{
  "model": "qwen/qwen3-0.6b-q8_0",
  "answers": {
    "urgent": {"type": "noul", "noul": 0.91},
    "owner": {
      "type": "choice",
      "choice": "infra",
      "probabilities": {"app": 0.08, "infra": 0.92}
    }
  },
  "usage": {"input_tokens": 81, "output_tokens": 4}
}

The probability fields are normalized decision weights derived from mean candidate token log-probabilities. They are useful for comparing the candidates in one question, but they are not calibrated probabilities of correctness. See Candidate score normalization for the exact calculation.

Error response

Invalid JSON, invalid request schemas, model failures, and other per-line failures have this shape:

{"error":{"code":"invalid_request","message":"..."}}

A failed line does not stop later input lines from running. Jet still exits nonzero after processing the stream so batch jobs can detect partial failure.

Inspect rendered prompts

export-prompts applies the model-independent Jet prompt templates without loading a model:

./target/release/jet export-prompts \
  --input requests.jsonl \
  --output -

This is useful when reviewing how state, instructions, and semantic candidates are presented before running inference.