Spring AI TypeSafe Documentation

Both layers ship separately. typesafe-java-sdk is the plain client (TypeSafeClient, question and answer model, retries, batches, typed exceptions) with no Spring AI dependency, so plain Java can use it. typesafe-spring-ai adds JevJudge, the advisors, and the RAG and tool-search integrations on Spring AI's own SPIs, and the starter auto-configures the client from spring.ai.typesafe.*. It is a spring-ai-community project, not a TypeSafe-published artifact.

No. The default suite runs offline with HTTP exercised through MockRestServiceServer, so ./mvnw clean verify needs no key and costs nothing. The *IT.java classes talk to the live API and are opt-in twice: behind the integration-tests profile and behind @EnabledIfEnvironmentVariable on TYPESAFE_API_KEY. One footgun remains, since the client falls back to TYPESAFE_BASE_URL and TYPESAFE_DEFAULT_MODEL.

Both judge through Jev, but only one retries. JevSelfRefineAdvisor judges an answer, feeds the defect back, and retries, which suits drafts. JevGuardrailAdvisor screens prompts and answers and does not retry, because an unsafe answer is not a draft. RAG also comes in two steps: JevDocumentFilter triages retrieved passages first, then JevDocumentReranker orders whatever survives.

That is the core pattern rather than a feature. Every question in a call is answered against the same state, so you compose atomic questions in code instead of writing one broad rubric prompt. Typed helpers return the pieces directly (noulValue, choiceValue, scoreValue, choice(...).confidence()), the client exposes retries and batch calls, and eight runnable demos sit in the examples module.

typesafe-java-sdk is the client with the question and answer model, retries, batches and typed exceptions and no Spring AI dependency. typesafe-spring-ai holds JevJudge, the advisors and the RAG and tool-search integrations on Spring AI's own SPIs. spring-ai-starter-typesafe adds properties and an auto-configured client bean, typesafe-bom manages versions, and examples holds eight runnable demos.

Java 17 or later. Spring AI 2.0.1 or later is needed for typesafe-spring-ai, while plain Java without Spring Boot needs only typesafe-java-sdk. The project is Apache 2.0 licensed and published on Maven Central.

Do not export TYPESAFE_BASE_URL or TYPESAFE_DEFAULT_MODEL while working on the SDK. TypeSafeClient.builder() falls back to both, so any throwaway main and the live integration tests will pick them up. Export TYPESAFE_BASE_URL only to run the ITs against a compatible server on purpose.