Struct EmbeddingEngine 

pub struct EmbeddingEngine { /* private fields */ }

Implementations

impl EmbeddingEngine

pub fn load(model_dir: &Path) -> Result<Self>

Load embedding model and vocabulary from a directory. Downloads model automatically from HuggingFace if not present.

pub fn load_model(base_dir: &Path, model_id: EmbeddingModel) -> Result<Self>

Load a specific embedding model from a directory.

pub fn load_default() -> Result<Self>

Load from default model directory (~/.lean-ctx/models/).

pub fn embed(&self, text: &str) -> Result<Vec<f32>>

Generate an embedding vector for a single text (document/code).

pub fn embed_query(&self, query: &str) -> Result<Vec<f32>>

Generate an embedding vector for a query string. Applies query-specific prefix if the model requires one.

pub fn embed_batch(&self, texts: &[&str]) -> Result<Vec<Vec<f32>>>

Generate embedding vectors for multiple texts using true batched ONNX inference. Sends a single [batch, max_seq_len] tensor through the model instead of batch separate calls — up to 50× faster on CPU for typical batch sizes (64–128) by leveraging matrix-matrix instead of matrix-vector operations inside the transformer.

pub fn dimensions(&self) -> usize

pub fn model_id(&self) -> &EmbeddingModel

pub fn model_name(&self) -> &str

pub fn model_directory() -> PathBuf

Resolve the model directory (respects LEAN_CTX_MODELS_DIR env).

pub fn is_available() -> bool

Check if the model files are present and loadable.

pub fn ensure_downloaded() -> Result<()>

Ensure the selected model’s files exist on disk, downloading them from HuggingFace if absent. Pure network/file IO — this never initializes the ONNX Runtime, so a failed bootstrap costs nothing in ORT teardown (the caller only loads ORT via shared_engine once the files are present). After Ok, is_available is guaranteed true; on a warm cache it short-circuits without touching the network.

The explicit index-build path uses this to cold-start a fresh machine. is_available() alone is only a file check, so the build path must call this to actually trigger the download instead of bailing (#545).