Struct EmbeddingService 

pub struct EmbeddingService { /* private fields */ }

Embedding service with Candle inference (no internal cache — use CachedProvider wrapper).

Implementations

impl EmbeddingService

pub fn new( model_dir: &Path, batch_size: usize, ) -> Result<EmbeddingService, CodememError>

Create a new embedding service, loading model from the given directory. Expects model.safetensors, config.json, and tokenizer.json in the directory.

pub fn default_model_dir() -> PathBuf

Get the default model directory path (~/.codemem/models/{MODEL_NAME}).

pub fn download_model(dest_dir: &Path) -> Result<PathBuf, CodememError>

Download the model from HuggingFace Hub to the given directory. Returns the directory path. No-ops if model already exists.

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

Embed a single text string. Returns a 768-dim L2-normalized vector.

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

Embed a batch of texts using a true batched forward pass.

Tokenizes all texts, pads to the longest sequence in each chunk, runs a single forward pass per chunk of up to batch_size texts, then performs mean pooling and L2 normalization on the batched output.

Trait Implementations

impl EmbeddingProvider for EmbeddingService

fn dimensions(&self) -> usize

Embedding vector dimensions.

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

Embed a single text string.

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

Embed a batch of texts (default: sequential).

fn name(&self) -> &str

Provider name for display.

fn cache_stats(&self) -> (usize, usize)

Cache statistics: (current_size, capacity). Returns (0, 0) if no cache.