Struct LazyFastembedProvider 

pub struct LazyFastembedProvider { /* private fields */ }

Lazy-loading embedding provider backed by FastembedProvider.

The constructor is instant — it sets initial state to Cached if model weights already exist on disk, or NotLoaded otherwise. The first call to embed triggers model download and ONNX session initialisation.

Thread-safe: concurrent callers block on a Condvar with a configurable timeout. After an idle period, the ONNX session is dropped to release memory while keeping weights on disk for fast reload.

Implementations

impl LazyFastembedProvider

pub fn new(model: EmbeddingModel, cache_dir: PathBuf, opts: LazyOpts) -> Self

Create a new lazy provider.

Returns instantly — no model download or ONNX initialisation.

pub fn new_with_loader( model: EmbeddingModel, cache_dir: PathBuf, opts: LazyOpts, load_fn: Arc<dyn Fn(EmbeddingModel, PathBuf) -> Result<FastembedProvider> + Send + Sync>, ) -> Self

Create a lazy provider with an injected model loader.

Production callers should use new; this constructor exists so tests can drive the panic/failure paths of ensure_model deterministically without real ONNX weights.

pub fn state(&self) -> ModelState

Current model lifecycle state.

pub fn is_ready(&self) -> bool

Whether the model is loaded and ready for inference.

Cheaper than state for hot-path polling — avoids cloning the Failed { message, .. } variant’s string.

pub fn reset(&self)

Clear a Failed state so the next ensure_model call retries the load, instead of returning the cached error forever.

No-op if the provider isn’t currently Failed. Resets to Cached if weights are already on disk, NotLoaded otherwise — mirroring the initial-state logic in new_with_loader.

Callers should check ModelState::is_panic before calling this: retrying after a panic-origin failure re-enters the same loader (and thus the same ort/ONNX init path) that just panicked, which may hit corrupted process-global state rather than a fresh, safe retry.

pub fn ensure_model(&self) -> Result<(), String>

Block until the model is ready for inference.

If the model is idle beyond the configured timeout, the ONNX session is evicted and reloaded from the on-disk cache. If the model is not yet loaded, this triggers download and initialisation on the calling thread. Concurrent callers wait on a Condvar for up to load_timeout_secs.

Trait Implementations

impl EmbeddingProvider for LazyFastembedProvider

fn dim(&self) -> usize

The dimensionality of the embedding vectors.

fn name(&self) -> &str

Provider name for display.

fn embed(&self, text: &str) -> Result<EmbeddingResult>

Embed a single text string.

fn embed_batch(&self, texts: &[&str]) -> Result<Vec<EmbeddingResult>>

Embed multiple texts in batch.