Struct StubEmbedder 

pub struct StubEmbedder { /* private fields */ }

Deterministic, dependency-free pseudo-embedder used in tests.

Hashes each word into a fixed-dim bucket (count-of-hash-buckets), then L2-normalizes the resulting vector. NOT semantic — texts that share words will be close; texts that don’t share words will be far. Good enough to exercise the hybrid-scoring code path without depending on a real ML model.

Default dimension is 8 (small enough to keep tests fast).

Implementations

impl StubEmbedder

pub const MODEL_ID: &'static str = "stub-d8"

pub const fn new() -> Self

pub const fn with_dim(dim: usize) -> Self

Trait Implementations

impl Clone for StubEmbedder

fn clone(&self) -> StubEmbedder

Returns a duplicate of the value.

fn clone_from(&mut self, source: &Self)

Performs copy-assignment from source.

impl Copy for StubEmbedder

impl Debug for StubEmbedder

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

Formats the value using the given formatter.

impl Default for StubEmbedder

fn default() -> Self

Returns the “default value” for a type.

impl Embedder for StubEmbedder

fn embed(&self, text: &str) -> Result<Vec<f32>, EmbedderError>

Compute an embedding for text. The returned vector MUST have length == self.dim(). Implementations may normalize.

fn model_id(&self) -> &str

Stable identifier for the model used. Stored alongside each embedding so retrieval can detect cross-model mismatches and skip the cosine blend rather than comparing apples-to-oranges.

fn dim(&self) -> usize

Embedding dimension. Used by the BLOB codec to validate stored vectors against the active model on retrieval.

fn is_noop(&self) -> bool

Convenience: true when this embedder is the production no-op. Callers can short-circuit the write/retrieval cosine path instead of allocating a vec only to discard it.

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

Embed many texts in as few backend calls as possible. The production fastembed backend runs them through ONNX in batched tensors (~10-40x faster than calling embed per text). The returned Vec is 1:1 with texts (same order, same length). The default impl loops embed, so non-batching embedders work unchanged. On any failure the whole batch errors — callers that want per-row resilience should fall back to embed per text.