Enum BenchEmbedder 

pub enum BenchEmbedder {
    BagOfTokens(ToyEmbedder),
    OnnxMiniLm {
        inner: Box<dyn Embedder>,
        model_id: String,
        dim: u32,
    },
}

Unified embedder used by crate::adapters::MnemAdapter.

The two flavours share a method surface (model(), dim(), embed_text()) so the adapter does not branch on the variant on every call. Construction is the only code path that picks one.

Variants

• BenchEmbedder::BagOfTokens - the always-compiled toy embedder. Selected by crate::EmbedderChoice::BagOfTokens.
• BenchEmbedder::OnnxMiniLm - real MiniLM-L6-v2 via mnem-embed-providers (gated on the onnx-minilm feature). Selected by crate::EmbedderChoice::OnnxMiniLm.

Variants

BagOfTokens(ToyEmbedder)

Toy hashed bag-of-tokens. Network-free, ~0.2 recall@5 on LongMemEval; ships as the offline / WASM-clean fallback.

OnnxMiniLm

Real all-MiniLM-L6-v2 via mnem-embed-providers. The concrete type is hidden behind a Box<dyn Embedder> so this crate stays agnostic to the underlying ORT session lifetime. model_id and dim are cached so the hot path (per-doc ingest, per-query retrieve) avoids vtable round-trips.

Fields

inner: Box<dyn Embedder>

Boxed provider implementing mnem_embed_providers::Embedder.

model_id: String

Cached fully-qualified model id (e.g. "onnx:all-MiniLM-L6-v2"). Identifies the vector lane keyed on the mnem retriever.

dim: u32

Cached output dimension (384 for MiniLM-L6-v2).

Implementations

impl BenchEmbedder

pub fn bag_of_tokens(dim: u32) -> Self

Construct the toy hashed bag-of-tokens embedder of dimension dim. Matches 0.1.0 behaviour.

pub fn onnx_minilm() -> Result<Self, Box<dyn Error>>

Construct the real ONNX MiniLM-L6-v2 embedder via mnem-embed-providers (onnx-bundled flavour). Lazy- downloads the model on first call (ORT + tokenizer + weights fetched into the HuggingFace cache; ~90MB).

Errors

Surfaces tokenizer / model-load / ORT-session failures from mnem-embed-providers verbatim as a Box<dyn Error>.

pub fn model(&self) -> &str

Fully-qualified model identifier. Stamped on every Embedding and used as the key the retriever’s vector lane resolves on, so two embedders with the same model() MUST produce vectors in the same semantic space.

pub fn dim(&self) -> u32

Output vector dimension.

pub fn embed_text(&self, text: &str) -> Result<Vec<f32>, Box<dyn Error>>

Embed a single string. Errors from the ONNX path (tokenizer, session.run) are surfaced as Box<dyn Error>. The toy path is infallible, so Result here is a small ergonomic tax we pay so the call site stays variant-blind.

Errors

Returns the underlying provider error verbatim for the ONNX flavour. The bag-of-tokens flavour cannot fail.

Trait Implementations

impl Debug for BenchEmbedder

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

Formats the value using the given formatter.