Local embedding model backend for Blazen using [fastembed].

This crate wraps the [fastembed] Rust crate (ONNX Runtime) to provide fully local, offline vector embeddings with no API keys required.

When used through blazen-llm with the fastembed feature flag, this crate's [FastEmbedModel] automatically implements blazen_llm::EmbeddingModel.

Target gating

As of ort-sys 2.0.0-rc.12, pyke dropped x86_64-apple-darwin from their prebuilt distribution matrix. On that target this crate compiles to a stub (see provider_stub) — the public API remains, but [FastEmbedModel::from_options] always returns [FastEmbedError::UnsupportedTarget]. Consumers should route through blazen-embed's facade, which picks blazen-embed-tract (pure-Rust ONNX) on those targets automatically.

Quick start

use blazen_embed_fastembed::{FastEmbedModel, FastEmbedOptions};

# async fn example() -> Result<(), Box<dyn std::error::Error>> {
let model = FastEmbedModel::from_options(FastEmbedOptions::default())?;
let response = model.embed(&["hello".into(), "world".into()]).await?;
assert_eq!(response.embeddings.len(), 2);
# Ok(())
# }