Struct StaticEmbedModel 

pub struct StaticEmbedModel { /* private fields */ }

Loaded Model2Vec static embedder.

Constructed via StaticEmbedModel::from_path. Use encode_query for a single text and encode_batch for many — the batch path is where the parallel-pool win lives.

Implementations

impl StaticEmbedModel

pub fn from_path(path: &Path, normalize_override: Option<bool>) -> Result<Self>

Load from a local directory containing tokenizer.json, model.safetensors, and config.json.

normalize_override lets callers force-enable or force-disable L2 normalization regardless of what config.json says. Pass None to honor the config.

pub fn from_bytes( tokenizer_bytes: &[u8], model_bytes: &[u8], config_bytes: &[u8], normalize_override: Option<bool>, ) -> Result<Self>

Load from in-memory bytes (e.g., for embedded resources or tests).

pub fn hidden_dim(&self) -> usize

Embedding dimension.

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

Encode a single text into a row vector.

Used at query time. The tokenization step is single-text so the nested-rayon trap doesn’t apply, but it’s a separate code path that avoids the unnecessary encode_batch_fast setup.

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

Encode a batch of texts.

Iterates over fixed-size sub-batches (BATCH_SIZE = 1024), each tokenized via encode_batch_fast (parallel internally inside tokenizers) and then mean-pooled via par_iter on the rayon pool. Calling one giant encode_batch_fast on a 250K-item corpus dominates wall time in Encoding allocation + internal chunk scheduling; the 1024-batch shape mirrors model2vec_rs’s internal default and measured noticeably faster on a 92K-file linux-source corpus.