Struct LocalEmbedder 

pub struct LocalEmbedder { /* private fields */ }

Implementations

impl LocalEmbedder

pub fn new(model: &str) -> Result<Self, String>

Build the embedder for the named model. Only all-MiniLM-L6-v2 is supported as the local backend (matches the prior fastembed surface).

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

Embed a batch of texts → one L2-normalized 384-dim vector each.

Internally sub-batches by a token budget so a single ONNX inference can never balloon peak RSS: the transient attention tensor scales with batch × heads × seq_len², so a batch that happens to contain many long (512-token) chunks would otherwise spike memory (~5 GB worst case at batch=64 × 512 tokens). We cap batch × max_len² per inference, which keeps short-chunk batches at full size (no throughput loss) while splitting long-chunk batches into smaller inferences. Output order and vectors are identical to embedding the whole input in one call.