@effect-uai/typesafe-ai

TypeSafe AI provider for @effect-uai/core. Serves the DecisionModel capability with Jev, a System One model that answers typed questions about an input with probabilities instead of generating text.

Experimental. This class of model is new and the shape may change.

Install

pnpm add @effect-uai/core @effect-uai/typesafe-ai effect

Wire it up

import { Config, Effect, Layer } from "effect"
import { FetchHttpClient } from "effect/http"
import { layer as jevLayer } from "@effect-uai/typesafe-ai/Jev"

const provider = Layer.unwrap(
  Effect.gen(function* () {
    const apiKey = yield* Config.Redacted("TYPESAFE_AI_API_KEY")
    return jevLayer({ apiKey })
  }),
)

const mainLayer = provider.pipe(Layer.provide(FetchHttpClient.layer))

One layer, two tags: Jev for its own reported confidence, DecisionModel for provider-portable code.

Decide

import { Schema } from "effect"
import * as Decision from "@effect-uai/core/Decision"
import { decide } from "@effect-uai/core/DecisionModel"

const triage = Decision.make({
  inputSchema: Schema.Struct({ message: Schema.String }),
  decisions: {
    department: Decision.classify({
      instructions: "Which team should handle this?",
      criteria: {
        billing: "Payments, invoicing, refunds",
        technical: "Bugs, outages, integrations",
      },
    }),
    severity: Decision.rate({
      instructions: "How severe is the reported issue?",
      criteria: ["Cosmetic", "Degraded", "Blocking"],
    }),
    urgent: Decision.probability({ instructions: "Does this convey urgency?" }),
  },
})

const program = Effect.gen(function* () {
  const { answers } = yield* decide(triage, {
    model: "jev-latest",
    input: { message: "Payouts have been failing for 3 days." },
  })

  answers.department.labels.billing // number
  answers.urgent.probability // number
  answers.severity.legend[Decision.topLevel(answers.severity)] // "Degraded"

  const leader = Decision.ranked(answers.department)[0]
  return leader === undefined || leader.probability < 0.5 ? "escalate" : leader.label
})

Every decision in a definition is answered in one call, and the encoded input is billed once, so batch them rather than calling per question.

Models

jev-latest (default), jev-preview, jev-1.13.0. 64k tokens per request, of which roughly 32k for the encoded input plus the longest single decision. Text only.

What the input schema is for

inputSchema encodes the input on the way out. A schema that omits a field keeps it from the model, which is the supported way to hold irrelevant or sensitive context out of a request. The vendor documents large irrelevant input as a distractor, so projecting deliberately is worth doing.

About the scores

Probabilities sum to 1 within floating-point error and are comparable within one answer. Jev's own confidence is available on the Jev tag; it is not on the common answer because its formula is undocumented and it is absent on probability decisions. Decision.confidence and Decision.margin are computed from the distribution, documented, and defined for every kind. Gate on the top probability and margin, not on confidence: normalised entropy shrinks as the label count grows, so a threshold tuned on two labels quietly demands far more mass on four.

Known weak spots per the vendor's own jaggedness notes: counting and arithmetic, date ordering, multi-hop reasoning, and rubric scores are weakly calibrated so expectedLevel should not be read back as a magnitude. Input is not treated as hostile, so an injected instruction in the input can steer an answer.

Errors

An input the schema cannot encode is InvalidRequest with param: "input". An answer whose shape does not match the decision that asked for it is GenerationFailed.

See also

• Decision models
• TypeSafe AI usage page
• @effect-uai/core/Decision