v0.2.0

Breaking

• Options.threads removed. Set the thread count once with configure_runtime(RuntimeConfig{n}) before the first open. All models share one process-wide ONNX Runtime pool, and after the first open a different value returns InvalidArgument.

Added

• find_model(name, id): looks up a model by its full upstream id or bare sentence-transformers name, ignoring case. Unknown names get a list of the supported models, and hosted API models (text-embedding-ada-002, text-embedding-3-*) get their own error.
• Seven new models, bringing the total to 17:
  • paraphrase-MiniLM-L6-v2
  • paraphrase-multilingual-MiniLM-L12-v2
  • paraphrase-multilingual-mpnet-base-v2
  • multi-qa-MiniLM-L6-cos-v1
  • multilingual-e5-base
  • snowflake-arctic-embed-m-v1.5
  • mxbai-embed-large-v1

Packaging

• New archives: ort-1.30.0-2 and tokenizers-0.23.2-3.
• Linux archives are built on manylinux_2_28.
• macOS archives target macOS 14.0.

Fixed

• Models whose ONNX output is named token_embeddings (sentence-transformers exports) now load.
• Registry generation no longer rejects normalize: false.

Full Changelog: v0.1.0...v0.2.0