v0.2.0
Highlights
Added ModernBERT-based embedding models and LightOnOCR, along with startup model preloading and explicit concurrency control in the asynchronous Python client.
Features
- Added GTE-ModernBERT-base, Snowflake Arctic Embed M v2.0, and IBM Granite English R2 embedding models, including the small variant.
- Added LightOnOCR-2-1B for OCR in the
transformers5bundle. - Added
max_concurrencytoSIEAsyncClientand Haystack-convention import aliases underhaystack_integrations. - Added anonymous usage telemetry, with opt-out through
SIE_TELEMETRY_DISABLED=trueorDO_NOT_TRACK=1.
Bug fixes
- Model-affinity routing can spill requests to other workers instead of becoming stuck, and rejected requests now contribute to autoscaling demand.
Breaking changes
- Worker startup no longer accepts
--modelto select models. Use--preloadorSIE_PRELOAD_MODELSto load models at startup; otherwise models load on demand.