Struct CandleEmbedder 

pub struct CandleEmbedder {
    device: Device,
    model: Arc<RwLock<Option<BertModel>>>,
    tokenizer: Arc<RwLock<Option<Tokenizer>>>,
    config: Arc<RwLock<Option<Config>>>,
    cache_dir: PathBuf,
    initialized: Arc<RwLock<bool>>,
}

GTE-small embedder using Candle.

Fields

device: Device

Device to run inference on (CPU or CUDA)

model: Arc<RwLock<Option<BertModel>>>

Loaded model

tokenizer: Arc<RwLock<Option<Tokenizer>>>

Tokenizer

config: Arc<RwLock<Option<Config>>>

Model configuration

cache_dir: PathBuf

Cache directory for models

initialized: Arc<RwLock<bool>>

Whether model is initialized

Implementations

impl CandleEmbedder

pub fn new(cache_dir: PathBuf) -> Self

Create a new CandleEmbedder.

pub fn with_device(cache_dir: PathBuf, device: Device) -> Self

Create with specific device.

pub async fn init(&self) -> Result<(), EmbedError>

Initialize the model (download if needed, load into memory).

fn mean_pooling( &self, token_embeddings: &Tensor, attention_mask: &Tensor, ) -> Result<Tensor, EmbedError>

Mean pooling with attention mask.

fn normalize(&self, embeddings: &Tensor) -> Result<Tensor, EmbedError>

L2 normalize embeddings.

async fn encode_batch( &self, texts: &[&str], normalize: bool, ) -> Result<Vec<EmbeddingOutput>, EmbedError>

Encode a batch of texts.

Trait Implementations

impl Embedder for CandleEmbedder

fn model_name(&self) -> &str

Model name/identifier.

fn dimension(&self) -> usize

Embedding dimension.

fn max_tokens(&self) -> usize

Maximum tokens per input.

fn modalities(&self) -> &[Modality]

Supported modalities.

fn embed_text<'life0, 'life1, 'life2, 'life3, 'async_trait>( &'life0 self, texts: &'life1 [&'life2 str], config: &'life3 EmbeddingConfig, ) -> Pin<Box<dyn Future<Output = Result<Vec<EmbeddingOutput>, EmbedError>> + Send + 'async_trait>>

where Self: 'async_trait, 'life0: 'async_trait, 'life1: 'async_trait, 'life2: 'async_trait, 'life3: 'async_trait,

Embed text content.

fn embed_query<'life0, 'life1, 'life2, 'async_trait>( &'life0 self, query: &'life1 str, config: &'life2 EmbeddingConfig, ) -> Pin<Box<dyn Future<Output = Result<EmbeddingOutput, EmbedError>> + Send + 'async_trait>>

where Self: 'async_trait, 'life0: 'async_trait, 'life1: 'async_trait, 'life2: 'async_trait,

Embed a query (may use different instruction).

fn embed_image<'life0, 'life1, 'life2, 'async_trait>( &'life0 self, _image_data: &'life1 [u8], _config: &'life2 EmbeddingConfig, ) -> Pin<Box<dyn Future<Output = Result<EmbeddingOutput, EmbedError>> + Send + 'async_trait>>

where 'life0: 'async_trait, 'life1: 'async_trait, 'life2: 'async_trait, Self: 'async_trait,

Embed image content.