Spring publishes a Java client that asks Jev before the chat model runs

Christian Tzolov published a Spring blog on September 21 about a community Java client for Jev. The repository is spring-ai-community/spring-ai-typesafe, Apache 2.0. It is a client of TypeSafe’s hosted API, plus Spring AI advisors that implement interfaces Spring already has: a judge, a guardrail, a document filter, a tool index, and an evaluator. The model is not wired up as a chat model. A call takes a state and a map of Noul, Choice, and Score questions.

The blog’s getting-started block depends on spring-ai-starter-typesafe and typesafe-spring-ai at version 0.1.0, and it says that version is on Maven Central. The README quick start on the commit we read still shows 0.1.0-SNAPSHOT, and its docs link points at latest-snapshot. Java 17 or later. The Spring AI module wants Spring AI 2.0.1 or later. A query to search.maven.org on September 22, for artifact spring-ai-starter-typesafe and for typesafe-java-sdk, returned no documents. The README badge points at central.sonatype.com for the SDK. We did not download a jar. Default tests talk to MockRestServiceServer. Live tests need a Maven profile and TYPESAFE_API_KEY, so an exported key does not turn an ordinary build into a billed one.

The blog says every number it quotes is from a live run. One support ticket, three questions, comes back as urgent 0.95, department billing at confidence 0.82, and a frustration Score of 1.1 on a rubric of Calm, Frustrated, Very angry. The distribution behind billing is 0.87 / 0.13 / 0.0, so the confidence is not the top probability. The same ticket text in the README is commented with different values: urgency 0.92, department technical, frustration 1.6, confidence still 0.82. Bare labels, with no description of what billing or technical means, drop that confidence to 0.60 in the blog. From Tzolov’s laptop, the one-question call had a median of 275 ms and the three-question call 310 ms, with 73 output tokens.

He also cites TypeSafe’s self-consistency cookbook: a 14-question call at $0.000043 and 111 ms, against $0.0018 and 1.8 seconds for claude-haiku-4-5 and about $0.033 and 11 to 14 seconds for the reasoning models in that note. The blog summarizes that comparison as 10 to 125 times faster and 22 to 805 times cheaper, and says it is TypeSafe’s benchmark and still needs checking. We did not rerun it. TypeSafe’s Master Customer Agreement section 2.3(f) forbids customers from publishing benchmarks of the Services.

JevJudge turns each criterion into a pass, a fail, or inconclusive. On a scripted answer that Paris is -455 degrees Celsius, helpfulness came back INCONCLUSIVE at 0.83 with confidence 0.44, plausibility FAILED at 0.02, and groundedness PASSED at 0.89. The default floor is 0.5, so an uncertain rubric does not fail the verdict by itself. JevSelfRefineAdvisor appends the failing criterion’s text and asks the chat model again, up to three times, rebuilding the prompt from the original each time. The log in the blog uses a scripted chat model that answers -125 C and then 15 C, because that demo also needs an Anthropic key. The blog says the Jev side of the log is a real call.

JevGuardrailAdvisor does not retry. Defaults treat jailbreak, physical harm, and illegal requests as BLOCK, and self-harm as SUPPORT, which refuses and points at help. A hazard probability above 0.70 applies that action. Between 0.35 and 0.70 the request is logged for review and still passes. Below 0.35 it passes. A severity above 2.0 turns a review into a block. The input check can return before the chat model is called. The output check is what catches an answer that crossed a line after an ordinary request. A Choice always names a winner, so the tool index asks a separate Noul for whether any tool applies at all.

Jev does not stream, so the advisors buffer. State has to be a string, an object, an array, or null. A bare number or boolean gets a 422. Reranking twenty documents is twenty calls, which is why the filter is meant to run first. We did not compile the examples.