Trait TextEmbedder

Source

pub trait TextEmbedder: Send + Sync {
    // Required method
    fn embed_texts(
        &self,
        texts: &[String],
        mode: EmbedMode,
    ) -> Result<Vec<Vec<f32>>>;

    // Provided methods
    fn embed_text(&self, text: &str, mode: EmbedMode) -> Result<Vec<f32>> { ... }
    fn model_id(&self) -> Option<&str> { ... }
    fn dimension(&self) -> Option<usize> { ... }
    fn capabilities(&self) -> TextEmbedderCapabilities { ... }
}

Expand description

Minimal interface for “text → dense vector” encoders (bi-encoder style).

This covers the common “sentence embedding” family: one vector per input string.

Important: there are multiple kinds of “embeddings” used in retrieval:

Dense sentence embeddings (this trait): one (d)-dim vector per string.
Dense token embeddings / late interaction (e.g. ColBERT): many vectors per string.
Sparse embeddings (e.g. SPLADE): a weighted sparse vector over vocabulary IDs.
Binary / quantized embeddings: compressed representations for ANN speed/memory.

We start with the dense-sentence contract because it’s the smallest stable surface that many parts of the workspace can share (iksh, chunking, retrieval prototypes).

Required Methods§

Source

fn embed_texts( &self, texts: &[String], mode: EmbedMode, ) -> Result<Vec<Vec<f32>>>

Embed texts into vectors.

Recommended invariant for backends that can afford it: L2-normalize outputs so cosine similarity equals dot product and downstream fusion is less brittle.

Provided Methods§

Source

fn embed_text(&self, text: &str, mode: EmbedMode) -> Result<Vec<f32>>

Convenience: embed a single text, avoiding the &[s.to_string()][0].clone() boilerplate.

Source

fn model_id(&self) -> Option<&str>

Optional: model identifier for debugging and provenance.

Source

fn dimension(&self) -> Option<usize>

Optional: embedding dimension as returned by this embedder.

Notes:

If you wrap an embedder with apply_output_dim(Some(d)), this should become Some(d).
Many remote backends can’t report this without a request; returning None is fine.

Source