Embedding services

The project provides a shared DJL-based text embedding starter (embedding-djl). It loads text and multimodal text models and exposes a DjlTextEmbeddingService bean to consuming modules (API and front-api).

Configuration

Configuration is driven by the embedding.* properties (DjlEmbeddingProperties):

• embedding.text-model-url / embedding.multimodal-model-url - remote identifiers for text embeddings (primary + fallback).
• embedding.vision-model-url / embedding.image-input-size - remote identifier and input size for image embeddings.
• embedding.fail-on-missing-model - when true, startup fails if neither text model nor multimodal fallback can be loaded.
• embedding.pooling-mode, embedding.engine - translator and engine tuning knobs (embeddings are always L2-normalized).

A health indicator (DjlEmbeddingHealthIndicator) publishes which model was loaded and its resolved path/URL when Spring Boot Actuator is present.

Usage

Add the embedding-djl module as a dependency and configure the properties above. The service loads remote DJL models directly and logs failures on startup when configured to fail fast.

Runtime packaging guidance

embedding-djl now performs a startup preflight (DjlTokenizersResourceValidator) that validates both DJL tokenizers metadata resources:

• native/lib/tokenizers.properties
• tokenizers-engine.properties

If these files are not visible to the runtime classloader, startup fails fast with diagnostics. For container/server deployments, prefer an exploded classpath launch (BOOT-INF/classes + BOOT-INF/lib/*) when executable nested jars hide JNI metadata resources.