TypeSafe SDK for Java

Community-maintained by Premo-Cloud and published to Maven Central as io.github.premo-cloud:typesafe-sdk, currently 0.6.0. The README is explicit that it is independent and not affiliated with, endorsed by or supported by TypeSafe AI. It follows the official Python and JavaScript SDK conventions so the three read alike, and depends only on Jackson.

Compile time, if you use the Asks form. An Ask names the key and, for a choice, the enum once, so reading an answer as the wrong type is a compile error and a misspelled label is a compile error too. Ask.choice(key, Dept.class, question) throws when the ask is created if a label is not a constant. Switch over the answer must cover every constant.

By default it retries twice after the first attempt on HTTP 408, 429 and 5xx, on connection failures and on timeouts, with exponential backoff from 500 ms capped at 5 s and 25 percent jitter, honoring Retry-After and retry-after-ms up to one minute. Retried attempts carry an X-TypeSafe-Retry-Count header, and a 429 still surfaces as TypeSafeRateLimitException with retryAfter().

No. The starter exposes a TypeSafeClient bean from typesafe.api-key and friends in application.properties, and backs off if you define your own. It also has a form for rules that arrive as data rather than code: CriteriaQuestionSet puts them in the state and generates one noul per entry pointing at its own criteria[i] path.

When the keys are only known at runtime, or when porting code from the Python and JavaScript SDKs. systemOne(state, questions) then takes Noul, Choice and Score keyed by ids you choose and reads answers back by the same ids. response.choice(key, Dept.class) reads a keyed enum choice as the enum and throws IllegalArgumentException naming the label when the enum lacks it, and nouls(), choices() and scores() return every answer of a kind by key.

No build() call. Every question type takes a configurer, so nested criteria read top to bottom in the style of the Elasticsearch and AWS Java clients: Noul.of takes whenTrue and whenFalse, Choice.of takes option(label, description), and Score.of takes ordered levels. Any description can be a plain string or a Criterion with what, notFor and examples, and prebuilt questions are plain records you can share across requests.

Rarely. Everything in one request runs in parallel on the server and shares a single round trip, and questions cannot see one another's answers, so you should ask independent questions together and include speculative ones your code may ignore. Start a second request only when an earlier answer is needed to build the next state, such as deciding what to fetch before you can phrase the follow-up.