Spring AI TypeSafe

Checked, not fully reviewed. We ran the default build, which is entirely offline — 286 tests, no failures — and drove the SDK against a stand-in. No Jev calls.

Four modules: a Java SDK with no Spring dependency, typesafe-spring-ai with the advisors, a Boot starter that reads spring.ai.typesafe.*, and a BOM. The advisors are the reason to use it: JevJudge for model-as-a-judge from atomic criteria, a self-refine loop, JevGuardrailAdvisor for input and output screening, a document filter and reranker for RAG, and JevToolIndex for tool selection that can say none apply. Live integration tests are opt-in twice — a Maven profile and an API key — so an exported key never turns a build into a billed one.

TypeSafe’s API docs say to back off and retry on 429 and on 529 Overloaded. We pointed the SDK at a server that answered 529 to everything: it made three attempts over 1.4 s and threw a typed TypeSafeOverloadedException, rather than returning anything. The guardrail advisor has no catch around its screening call, so when Jev is down a guarded turn fails instead of reaching the model — the safe direction for a guardrail, though it is not one of its ten tests. Self-harm signals route to a support message rather than a refusal, which is a considered touch.

One documented sharp edge: TypeSafeClient.builder() falls back to TYPESAFE_BASE_URL and TYPESAFE_DEFAULT_MODEL, so an exported variable redirects every default-built client. The README warns about it in bold. Apache-2.0, 47 stars.

See how it compares with other tools in Best Jev tools, tested hands-on.

Review updated Oct 4, 2026. Numbers quoted from the project are its author's own; we don't publish our own measurements of Jev.