Struct OnnxEmbeddingProvider 

pub struct OnnxEmbeddingProvider { /* private fields */ }

A MiniLM-class sentence-embedding model, in this process.

Implementations

impl OnnxEmbeddingProvider

pub async fn load(config: OnnxProviderConfig) -> Result<Self, OnnxLoadError>

Load the artifacts and open a session.

The hang this guards against

ort 2.0.0-rc.12 hangs forever instead of erroring when the library at ORT_DYLIB_PATH cannot be loaded — measured in this repository against a nonexistent path, where neither Session::builder nor ort::init_from returns in 45 seconds. No pre-flight validation exists; every entry point funnels through the same lazy init. Since the hang cannot be interrupted from inside, the blocked thread is abandoned: a raw std::thread rather than spawn_blocking, because Tokio joins blocking threads at shutdown and a wedged one would hang that too. A caller receiving OnnxLoadError::RuntimeHung has leaked one thread and must terminate the process rather than retry.

See gears/file-parser/file-parser/src/gear.rs for the same mitigation and the measurements behind it.

Errors

Unreadable artifacts, an unloadable tokenizer, a session that refuses to open, or the runtime hang above.

Trait Implementations

impl EmbeddingProviderV1 for OnnxEmbeddingProvider

fn embedding_space(&self) -> &EmbeddingSpaceId

Model artifact, tokenizer artifact, preprocessing and pooling configuration — not just a dimension.

fn dimension(&self) -> u32

fn embed<'life0, 'async_trait>( &'life0 self, req: EmbedRequest, ) -> Pin<Box<dyn Future<Output = Result<EmbedResponse, EmbeddingProviderError>> + Send + 'async_trait>>

where Self: 'async_trait, 'life0: 'async_trait,

fn health<'life0, 'async_trait>( &'life0 self, ) -> Pin<Box<dyn Future<Output = Result<(), EmbeddingProviderError>> + Send + 'async_trait>>

where Self: 'async_trait, 'life0: 'async_trait,