v0.5.0

Highlights

Self-hosted worker pools can now serve multiple bundles with separate replica limits, making it possible to scale embedding and generation workloads independently on the same machine profile.

Features

• Added Granite Guardian 3.0 2B for content-safety verdicts, with a configurable verdict threshold. Added SQLCoder-7B-2 for completion-based SQL generation using its native prompt format.
• Added configurable code, sql, and guard model aliases and exposed matching capabilities in model metadata. The default sql alias uses Qwen3-4B-Instruct-2507, not SQLCoder.
• Added a separate gateway metrics listener for Prometheus scraping without opening inference endpoints to unauthenticated access.

Bug fixes

• Florence-2 extraction now honors the supplied instruction.
• Guard-model verdict handling now keeps returned log probabilities consistent and rejects unsupported multi-candidate sampling.
• Helm rejects missing or invalid per-bundle replica limits during rendering.

Breaking changes

• Move each pool's bundle, minReplicas, maxReplicas, extraEnv, and imageBundle settings into workers.pools.<pool>.bundles.<bundle>. The bundle name becomes the map key; workers.common.bundle is removed.
• Worker resource names change from worker-<pool> to worker-<pool>-<bundle>. Upgrades must explicitly remove obsolete StatefulSets, ScaledObjects, PodDisruptionBudgets, and image-prepull DaemonSets so old resources do not interfere with scaling or node drains.