Zio TypeSafe AI

As data. Instructions and every criteria entry accept any zio.schema.Schema value, so a case class, Map, collection or Option.empty for JSON null, and Content wraps structured descriptions per option, per level, or on each Noul side. Schemas, taxonomies, database rows and boundary examples therefore go over the wire as structured values instead of being interpolated into prose.

You get a compile-valid Left rather than a failure surfacing later on the wire. Jev accepts 1 to 255 choice options and 2 to 10 score levels, so constructing outside that range returns an Either at build time. Likewise every noul, confidence and probability is a Probability, a Double provably within [0.0, 1.0], rather than an unconstrained number.

They compose as middleware using the same @@ and ++ shape as ZIO HTTP. Middleware.make wraps the exchange effect, and the example retries rate limits by catching Error.RateLimit around next, while Middleware.logging records requests, canonical responses, failures and latency. Observed responses are decoded and re-encoded canonically rather than kept as byte-exact HTTP text.

Logging is opt-in and never includes headers or bearer tokens, but state and model output can still be sensitive, so it stays off until you configure it. The client is a ZIO HTTP Client layer: provide Client.default together with TypeSafeAI.Client.live, or build one from an explicit ApiKey and ModelId.JevLatest.

Add the Maven coordinate com.jamesward by zio-typesafe-ai to libraryDependencies; it is Scala 3 built on ZIO HTTP's Client. TypeSafeAI.Client.live reads the same environment variables as the official SDKs: TYPESAFE_API_KEY is required and TYPESAFE_DEFAULT_MODEL is optional, defaulting to jev-latest. You can also build the layer explicitly from a TypeSafeAI.ApiKey and TypeSafeAI.ModelId.JevLatest. It is Apache-2.0 licensed.

You bundle questions in a NamedTuple passed to TypeSafeAI.ask; the keys become both the wire question ids and the compile-time field names of the typed result. Ask for isUrgent and department and you get back a value with .isUrgent typed as a Probability and .department typed as a ChoiceAnswer. Passing anything other than a non-empty NamedTuple of Questions is a compile error, and state plus instructions are generic in a Schema type parameter.

.answers returns just the typed NamedTuple. .run wraps it in a Result that also carries the responding model and token usage. Both accept the same questions, and you can override the model per call, for example with .model(TypeSafeAI.ModelId("jev-2025-11")). Every noul, confidence and probability is a Probability, a Double provably within [0.0, 1.0], so out-of-range values are not representable.