Struct FastembedBackend 

pub struct FastembedBackend { /* private fields */ }

ONNX-backed embedding via fastembed-rs. Loads the model at construction; inference runs on a tokio blocking thread per batch.

The inner TextEmbedding is held behind a mutex, which serializes .embed() calls across concurrent users of the same FastembedBackend. For schema-RAG (low QPS, small batches) this is fine. For bulk user-data embedding (Phase C+) we may want a pool of TextEmbedding instances or a lock-free inference path — revisit when that workload materializes.

Implementations

impl FastembedBackend

pub async fn new( model_id: &str, model_path: Option<&str>, ) -> Result<Self, EmbedError>

model_id is the short name (e.g., "bge-small-en-v1.5"). model_path optionally points at a pre-downloaded ONNX dir for air-gapped deployments (env KYMA_EMBED_MODEL_PATH).

Trait Implementations

impl Debug for FastembedBackend

fn fmt(&self, f: &mut Formatter<'_>) -> Result

Formats the value using the given formatter.

impl EmbeddingBackend for FastembedBackend

fn id(&self) -> &str

Stable identifier, e.g. "fastembed/bge-small-en-v1.5" or "openai/text-embedding-3-small". Mixing IDs across writes is a correctness bug (distance-space mismatch) — the engine tags every vector column with its model_id.

fn dimension(&self) -> u16

Output dimension. Must be stable for the lifetime of the instance.

fn embed<'life0, 'life1, 'async_trait>( &'life0 self, texts: &'life1 [String], ) -> Pin<Box<dyn Future<Output = Result<Vec<Vec<f32>>, EmbedError>> + Send + 'async_trait>>

where Self: 'async_trait, 'life0: 'async_trait, 'life1: 'async_trait,

Compute embeddings for a batch. Order of outputs matches inputs.