@jsilvanus/seedeer@0.3.0
Minor Changes
-
a3b42c0: Update default models to more contemporary alternatives:
embedeer: default embedding model is nowonnx-community/gte-multilingual-base(768-dim, multilingual GTE) instead of the 2021-eraXenova/all-MiniLM-L6-v2. Also fixes an existing bug where theEmbedderconstructor's fallback model name (nomic-embed-text) was not a valid Hugging Face repo id.seedeer:Detector's default model is nowonnx-community/rtdetr_r18vd(RT-DETR) instead ofXenova/yolos-tiny.VisualEmbedderexamples/benchmarks now use DINOv3 (onnx-community/dinov3-vit{s,b}16-pretrain-lvd1689m-ONNX) instead of DINOv2.
These are drop-in replacements using the same pipeline/model-class APIs.
Captioner(BLIP/Florence-2 alternatives aren't simple drop-ins for theimage-to-textpipeline) andJointEmbedder(switching CLIP to SigLIP requires different model classes) were left unchanged — flagged separately for a follow-up.
Patch Changes
-
0327647: Fix markdown links pointing to old standalone-repo GitHub orgs; point them at the monorepo (
jsilvanus/deer) instead. -
f2a02d5: No functional change — documents in TODO.md the deferred
Captioner(BLIP/Florence-2) andJointEmbedder(SigLIP) model upgrades that were left out of the contemporary-defaults pass. -
6499111: Unify GPU provider resolution across all packages via nudeer.
nudeer: Export sophisticated
resolveProvider(device, provider)with CUDA library verification, ldconfig caching, and detailed error messages. Replaces simpleresolveDeviceto provide production-quality GPU detection.embedeer & seedeer: Refactor to import
resolveProviderfrom nudeer. All packages now use the same CUDA detection logic with library verification and helpful installation instructions. No public API changes.Consolidates provider detection code, eliminating 3 duplicate implementations.
-
Updated dependencies [0327647]
-
Updated dependencies [6499111]
- @jsilvanus/nudeer@0.2.0