Skip to main content

Decisions

How to ask typed questions about an input and get a choice, a score or a probability back, from a System One model: one that decides without writing text, in well under a second.

import { choice, noul, score } from "@leuria/client"

const { answers } = await ai.decide(
  { ticket: "My flight was cancelled. Can I get a refund?" },
  {
    refund: noul("Does `ticket` ask for a refund?"),
    kind: choice("What is asked in `ticket`?", { refund: "money back", rebooking: "another flight", info: "a question" }),
    anger: score("How angry is `ticket`?", ["calm", "upset", "furious"]),
  },
)
answers.refund.noul        // 0 (no) to 1 (yes)
answers.kind.choice        // "refund" | "rebooking" | "info", with answers.kind.probabilities and .confidence
answers.anger.score        // 0 to 2, possibly between levels

Use it where a chat is too slow, too costly or too free-form: routing a message, sorting or scoring a list, a guardrail before an action, deciding whether something deserves a notification.

ai.decide()​

ai.decide(state, questions, options) asks the first ready provider that offers decide. Today that is the visitor's own model through Leuria: a decision model on their computer (LM Studio and Ollama both serve nimble, tev1… at /v1/systemone), or a service they set up: TypeSafe's Jev with their own key, or any other speaking the same API. The page holds no key and pays nothing.

  • state: what the questions are about. A string, or an object whose fields the questions name in backticks.
  • questions: an object of question id → question. Three kinds, with helpers that keep the answer types:
    • noul(instructions, criteria?): yes or no. The answer is { noul }, a probability. criteria may describe what true and false mean.
    • choice(instructions, criteria): one option among criteria (option → description, or null). The answer is { choice, probabilities, confidence }.
    • score(instructions, criteria): a level on criteria, 2 to 10 descriptions from low to high. The answer is { score, legend, probabilities, confidence }.
  • options.model: a model of the service (its latest by default). Ignored when the visitor decides with a model on their computer: that one is their pick. options.provider, options.signal: as for chat.

The result is { answers, model, provider, usage? }. With nothing ready, it rejects with NoProviderError. A question the service refuses (too many options, a missing field) rejects with its message.

The request and the answers are System One's own shapes, passed through: what their documentation says about writing questions applies as it is.

Who decides​

  • ai.getState().decider is the provider and model decide() would use now.
  • ai.getState().decidePending is a decider ahead of it that needs a click (connect Leuria).

Declare it, so Leuria tells the visitor when they connect: leuria({ needs: { decide: true } }). Leuria uses the first decision model found on the visitor's computer, or the one they chose (Your AIs › Decisions): a local model, TypeSafe with their key, another service, or off. The visitor can also give one site its own choice (Websites › the site › Decision model). Without one, leuria() has no decide capability and the page sees no decider. Where the request goes depends on that choice (their computer, or a service), so localOnly does not apply; the provider's locality stays device.