inference4j

inference4j wraps ONNX Runtime with Java APIs for local text, vision and audio inference. Its README requires Java 17 or later, documents Gradle and Maven dependencies, and describes automatic model downloads with a local cache. Local inference needs no model-service API key; initial model resolution can require network access. ONNX Runtime requires the documented JVM native-access permission.

An October 14 source commit adds a DeBERTa prompt-injection detector for user input and retrieved documents. It uses overlapping windows for long text and exposes a configurable threshold. At review, the latest packaged release remains v0.13.1, published October 12; the new detector requires the later source code. The default model is English-focused, and the documentation calls detection one layer of defense rather than comprehensive protection.