Jev

@typia/jev converts the questions of typia.llm.evaluation<T>() to the wire format of Jev , TypeSafe’s System One model. Jev answers typed questions with calibrated probabilities, and never returns a value outside the declared options.

typia.llm.evaluation<T>() emits the provider-neutral questions of Vercel AI SDK’s experimental_evaluate(), which spells a yes/no question "boolean". Jev’s own wire format spells it "noul", and that format is shared by TypeSafe’s API and SDK and by OpenRouter’s Decisions API. toJevQuestions() renames exactly that question type; choice and score questions pass through unchanged.

npm install typia @typia/jev @typesafe-ai/sdk
npx typia setup

Usage

import { TypeSafeClient } from "@typesafe-ai/sdk";
import { toJevQuestions } from "@typia/jev";
import typia from "typia";
 
interface ITriage {
  /** Is the ticket urgent? */
  urgent: boolean;
 
  /** Which team owns the ticket? */
  team: "billing" | "technical";
}
 
const evaluation = typia.llm.evaluation<ITriage>();
const client = new TypeSafeClient(); // reads TYPESAFE_API_KEY
const { answers } = await client.systemOne({
  model: "jev-1.13.0",
  state: "The payment page crashes for every customer.",
  questions: toJevQuestions(evaluation.questions),
});
const result = evaluation.validate(answers);

The answers need no conversion: validate() accepts Jev’s native { type: "noul", noul } answer as it is, and folds every answer back into ITriage by the conversion rules of typia.llm.evaluation<T>(). The raw answers keep what ITriage has no room for, such as the probability of every option.

Pin a versioned model once you tune probability thresholds, because TypeSafe’s aliases such as jev-latest move when a new release ships.

Wire format

toJevQuestions() returns a new map keyed like its input:

The returned questions type-check against @typesafe-ai/sdk’s Question, so they can be passed to systemOne() directly.