Expand description
Pure-Rust ONNX inference backend for Blazen embeddings via tract-onnx.
Mirrors the public API of blazen-embed-fastembed so the two backends are
swappable via cfg gating in blazen-llm. Exists because fastembed/ort
require Microsoft’s prebuilt ONNX Runtime binaries which are not published
for several target triples (notably *-unknown-linux-musl and wasm32-*).
§wasm32 support
On wasm32-* targets the native provider module is compiled out
because its from_options constructor relies on tokio runtime primitives
and on blazen-model-cache’s HuggingFace Hub downloader, neither of
which compiles to wasm32. In its place, [wasm_provider] exposes
[WasmTractEmbedModel] which downloads ONNX weights and the tokenizer via
web_sys::fetch and runs inference through the same tract-onnx pipeline.
Re-exports§
pub use options::TractOptions;pub use provider::TractEmbedModel;pub use provider::TractError;pub use provider::TractResponse;
Modules§
- options
- Options for
TractEmbedModel. - provider
- The
TractEmbedModeltype providing local embeddings viatract-onnx.