betto_inferencing 0.1.0-dev.3

ONNX Runtime inference and embedding models for dense text retrieval.

Changelog

0.1.0-dev.3

Features

• multilingual-e5-small registered in ModelCatalog and validated — this package's first cross-lingual embedding model, ~100 languages, 384 dimensions (same as bge-small-en-v1.5, so no SQ8/index-format change for consumers switching models). Uses the plain fp32 model.onnx export (not a quantized or GPU-oriented variant — see the ModelCatalog doc comment for the trade-off analysis) and XlmRobertaTokenizer for tokenization. Real checksums verified via the new tool/register_model.dart.
• EmbeddingKind — new document / query enum, and a corresponding kind parameter on EmbeddingModel.embed() (default EmbeddingKind.document, source-compatible for existing callers). OnnxEmbeddingModel.embed() uses it to apply ModelSpec.meta's 'documentPrefix' / 'queryPrefix' when the loaded model defines them — multilingual-e5-small requires a mandatory "passage: " / "query: " prefix per its model card; bge-small-en-v1.5 has neither key, so its behaviour is byte-for-byte unchanged.
• ModelTokenizer — new shared interface (encode(String) -> TokenizerOutput) implemented by both BertTokenizer and XlmRobertaTokenizer. OnnxEmbeddingModel.load() selects the concrete implementation via a new ModelSpec.meta['tokenizerFamily'] key ('bert' or 'xlmr', added explicitly to bge-small-en-v1.5's own entry too) rather than a compile-time fork, so a third tokenizer family is additive.
• tool/register_model.dart — new one-off dev tool that downloads a ModelCatalog entry's asset files from the exact URL ModelSpec uses at runtime and computes their SHA-256 digests via package:crypto, so a registered checksum is guaranteed to match what a real download verifies.

Changes

• placeholder-model replaces the previous bge-m3-v1.0 stub entry in ModelCatalog. The old entry had placeholder all-zero checksums and could never actually be downloaded or validated — a registered model that would silently fail with a confusing checksum-mismatch error rather than a clear "not supported" message. placeholder-model is a permanent, deliberately validated: false test fixture (non-resolvable example.invalid URLs) that exists solely to give tests a stable, always-unvalidated registered id to assert catalog gating behaviour against. Registering bge-m3 properly (a real 1024-dimensional multilingual model) is deferred future work — its ONNX export exceeds the 2 GB single-file limit and needs ModelSpec/ModelDownloader support for a split model.onnx + model.onnx_data layout first.

0.1.0-dev.2

Features

• XlmRobertaTokenizer — XLM-RoBERTa-family SentencePiece/Unigram tokenizer (e.g. multilingual-e5-small), returning the same TokenizerOutput type BertTokenizer uses. Composes a from-scratch CharsmapTrie (a Darts double-array trie reader for SentencePiece's precompiled_charsmap normalizer) with dart_sentencepiece_tokenizer's public API, working around two defects in that package's HuggingFace tokenizer.json loading path — see README.md for details and NOTICE for third-party attribution.
• tool/generate_xlmr_parity_corpus.dart and test/fixtures/xlmr_parity_corpus.json — a 58-language (+3 edge case) byte-exact tokenizer parity corpus, extracted from the NLTK UDHR corpus (same source betto_lang_detector uses) and annotated with real AutoTokenizer token ids. Gated by a new test in the test-macos integration suite, additive to the smaller 11-entry gate already in place.

Fixes

• XlmRobertaTokenizer — empty-string input no longer produces a spurious extra token. _metaspace was unconditionally adding a dummy-prefix space before replacing spaces with ▁, so "" became "▁" — itself a valid standalone vocabulary piece — yielding [<s>, ▁, </s>] instead of the real AutoTokenizer's [<s>, </s>]. Found by the new 61-entry parity corpus above.

0.1.0-dev.1

Initial development release providing ONNX Runtime inference and embedding models for dense text retrieval on native platforms (macOS, Linux, Windows, Android, iOS).

Features

• EmbeddingModel — abstract interface for text-to-vector embedding, decoupling consumers from any specific inference backend.
• OnnxEmbeddingModel — ONNX Runtime implementation backed by BGE Small En v1.5, delivering dense embeddings suitable for semantic search and retrieval tasks.
• BertTokenizer / TokenizerOutput — BERT WordPiece tokenizer with configurable word segmentation (default RegExpTokenizer; drop-in IcuTokenizer support via package:betto_icu).
• ModelCatalog — allowlist provider that gates model use behind download-on-demand via ModelDownloader from betto_onnxrt. Registered models: bge-small-en-v1.5 (validated), bge-m3-v1.0 (registered, not yet validated — throws UnsupportedError).
• quantise / dequantise — SQ8 scalar quantization helpers for compact vector storage and retrieval.
• Re-exports DownloadProgress, ModelDownloader, ModelFile, ModelSpec, and ResolvedModel from betto_onnxrt as part of the stable public API.