Primitives

Noul, Choice and Score. When to use which, and what comes back.

Follows docs.typesafe.ai/primitives. Every tool in this package is built from these three; jev_ask exposes them directly.

primitive	the question shape	comes back as	reach for it when
Noul	yes or no	noul, the probability of yes	one fact about the state: does it ask for a refund, is it urgent, does the claim appear in the source
Choice	pick one of the options you list, up to 255	choice, confidence, probabilities over every option	classification, routing, picking a candidate, extracting a value from a closed set
Score	place the state on an ordered scale of 2 to 10 levels you describe	score (probability-weighted position, index 0 the first level), confidence, legend, probabilities	severity, frustration, quality, fit: anything with an order

Writing them

Each question has instructions, the question the model reads, and optional criteria, which say what each answer means. For a Noul the criteria are true and false descriptions; for a Choice they map each option to when it applies; for a Score they are the ordered level descriptions. In jev_ask a question is a plain dict:

questions = {
    "urgent": {"type": "noul", "instructions": "Does this convey urgency?",
               "criteria": {"true": "the writer needs action now", "false": "it can wait"}},
    "team": {"type": "choice", "instructions": "Which team should take this?",
             "criteria": {"billing": "charges and payouts", "technical": "bugs and outages"}},
    "severity": {"type": "score", "instructions": "How bad is the impact?",
                 "criteria": ["cosmetic", "degraded", "blocked"]},
}

A bare string is a Noul. A list of specs is named q1..qN. Instructions and criteria may be JSON objects instead of strings when the question itself has structure; see State.

Reading them

A Noul is a probability, not a verdict: compare it to a threshold your code owns. A Choice's choice is the top option; look at probabilities when the runner-up matters, and add a none_of_these option when nothing may fit (the routing tools do this for you). A Score's score is a position on your scale, and legend names the levels, so read "1.04, just above frustrated" rather than "1.04". Noul answers carry no confidence; the probability is the whole answer.

The constraints

Options in a Choice: 255. Levels in a Score: 2 to 10, in order, index 0 first. Question keys are yours and are not shown to the model. All of these are checked by the tools before a request is sent; a violation is an error naming the limit, not a clamped request.

Edit page