CaseHub Neural Text — The Qualifier That Couldn’t Share Its Name
Date: 2026-06-08 Type: phase-update
What I was trying to achieve: CDI wiring for inference models
C4 shipped the task adapters. C6 needed the SPLADE sparse embedder. C5 needed to wire everything into Quarkus CDI so downstream consumers can inject models by name. The issue description called for @InferenceModel as the qualifier annotation. That lasted about thirty seconds of actual design work.
The naming collision nobody thought about
The SPI interface is io.casehub.inference.InferenceModel. The qualifier would be io.casehub.inference.quarkus.InferenceModel. Different packages, same simple name. Java’s import system can’t handle that — you can’t import both. Consumer code would look like:
@Inject
@io.casehub.inference.quarkus.InferenceModel("nli")
io.casehub.inference.InferenceModel nliModel;
Fully qualified everything. Ugly, error-prone, and a daily annoyance for anyone writing injection points. The fix was obvious once the problem was visible: call it @Inference instead. Short, unambiguous, lives in the quarkus package where context makes it clear. The issue title was wrong and the design is better for it.
CDI without a Quarkus extension
The full Quarkus extension pattern — deployment module, build-time synthetic beans, the works — is the proper way to do named resource injection. It’s also gated by the native image prototype, and the inference modules are JVM-only until that lands.
The alternative: a single InferenceModelProducer with @DefaultBean, a @Produces method that inspects the InjectionPoint for the @Inference qualifier value, and a ConcurrentHashMap for lazy model creation. Configuration via @ConfigMapping with a Map<String, ModelProperties>. The producer reads the qualifier name, looks up the config, creates an OnnxInferenceModel on first access, and closes everything on shutdown.
@Produces @DefaultBean @Inference("") @Dependent
InferenceModel produce(InjectionPoint ip) {
String name = ip.getAnnotated().getAnnotation(Inference.class).value();
return models.computeIfAbsent(name, this::createModel);
}
The @Nonbinding value on @Inference means CDI treats all @Inference qualifiers as the same bean regardless of the name — the InjectionPoint inspection resolves the specific model at runtime. @DefaultBean means any test that provides its own @Produces @Inference("") InferenceModel method wins automatically. The test override uses @Alternative enabled via QuarkusTestProfile.getEnabledAlternatives() — production producer is invisible, stub returns InMemoryInferenceModel, no config needed.
The JNI classloader surprise
The inference-quarkus module had a pre-existing plain JUnit test that loads ONNX Runtime JNI directly. Adding @QuarkusTest classes broke it — UnsatisfiedLinkError: Native Library already loaded in another classloader. The symptom looks like a library packaging problem. The cause is surefire’s fork reuse: the @QuarkusTest loads the JNI library via the Quarkus augmented classloader, and the plain JUnit test tries to reload it via the system classloader in the same JVM fork. JNI refuses. The fix is reuseForks=false — each test class gets its own JVM.
SPLADE: the simplest adapter
SparseEmbedder follows the same pattern as the C4 task adapters — takes an InferenceModel, applies post-processing, returns a typed result. The post-processing is three lines: ReLU, log1p, threshold filter. The output is Map<Integer, Float> — vocabulary token indices to weights, suitable for direct Qdrant sparse vector upsert. The whole class is under 70 lines. Sometimes the right abstraction is just a function with a name.