Triage a message

Route a support message without a chat: a decision model answers typed questions with a choice, a score and a probability, in well under a second. This needs a decision model: one on your computer through Leuria (nimble or tev1 in LM Studio or Ollama), or a System One service you set up in Leuria (TypeSafe with your key).

This is the classic Jev example: which team, how frustrated, how urgent, then act on the numbers.

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

const message = "Hi, I've been trying to connect my Stripe account for 3 days and the integration keeps failing. I'm losing sales. Please help ASAP."

const { answers, model } = await ai.decide(
  { message },
  {
    department: choice("Which team should handle `message`?", {
      billing: "Payment or subscription issues",
      technical: "Bugs or integration problems",
      sales: "Pricing or account questions",
    }),
    frustration: score("How frustrated does the writer of `message` appear?", [
      "Calm, just stating facts",
      "Frustrated but civil",
      "Very angry, strong language",
    ]),
    is_urgent: noul("`message` conveys urgency or time-sensitivity"),
  },
  { signal },
)

// Numbers your code can act on, not prose to parse.
const { department, frustration, is_urgent } = answers
const queue = is_urgent.noul > 0.7 ? `${department.choice}-urgent` : department.choice
const escalate = frustration.score >= 1.5 || department.confidence < 0.5

print({
  model,
  queue,
  escalate,
  department: `${department.choice} (${Math.round(department.confidence * 100)}% sure)`,
  frustration: `${frustration.score.toFixed(1)} of ${Object.keys(frustration.legend).length - 1}`,
  urgent: `${Math.round(is_urgent.noul * 100)}%`,
})

How it works

• ai.decide(state, questions) sends the state and the questions to the first provider that offers decide, and returns one answer per question, typed by the question's kind. Questions name fields of state in backticks.
• choice answers with the most likely option, the probabilities of every option, and a confidence; score with a level that can fall between two descriptions, and the legend; noul with a probability.
• The thresholds are yours: is_urgent.noul > 0.7, confidence < 0.5. Calibrated numbers are what make a decision model worth it over a chat answer.
• Nothing ready: NoProviderError, with why. Ask the visitor to connect Leuria, or to load a decision model. See Decisions.