Struct StaticEmbedder 

pub struct StaticEmbedder { /* private fields */ }

A loaded model2vec-format model. Cheap to share behind an Arc; embed takes &self.

Implementations

impl StaticEmbedder

pub fn load(paths: &ModelPaths, manifest: &ArtefactManifest) -> Result<Self>

Loads and verifies a model from local files. Fails, never warns, when either file’s blake3 digest differs from the manifest, when the files do not parse, when the matrix is not [vocab, dim], contains a non-finite value, or has fewer rows than the tokenizer has ids.

pub fn load_with( paths: &ModelPaths, manifest: &ArtefactManifest, opts: LoadOptions, ) -> Result<Self>

pub fn describe( paths: &ModelPaths, manifest: &ArtefactManifest, ) -> Result<Description>

What a loaded model would report as its profile and dimension, without loading it.

For status, doctor and a scan with nothing to embed (2026-09-25: each of them parsed the tokenizer and read the matrix for these two values, 2.6 s). Both files are verified against the manifest exactly as load does, the weights through the same staleness record; only the tokenizer is not parsed and the matrix not read beyond its header. A tokenizer that does not parse is therefore found by the first recall, not here.

pub fn embed_one(&self, text: &str) -> Result<Embedding>

Embeds one input. Errors only on a tokenizer failure, which no ordinary text causes.

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

Embeds a batch, in input order, splitting large batches across rayon’s pool.

pub fn vocab_rows(&self) -> usize

pub fn unk_id(&self) -> Option<u32>

pub fn manifest(&self) -> &ArtefactManifest

pub fn info(&self) -> ModelInfo

Trait Implementations

impl Debug for StaticEmbedder

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

Formats the value using the given formatter.

impl Embedder for StaticEmbedder

fn dim(&self) -> usize

Vector dimension. Constant for the life of a profile.

fn profile_id(&self) -> &str

Identifies model, dimension and pooling together. Stored alongside the vectors so a changed model is detected instead of silently degrading search (SPEC §5).

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

L2-normalised vectors, one per input, in input order.