Struct BgeM3Embedder 

pub struct BgeM3Embedder { /* private fields */ }

BGE-M3 multi-function embedder via Ollama.

Produces three representations from a single Ollama model call:

1. Dense vector — direct from Ollama’s /api/embed endpoint.
2. Sparse weights — derived from the BGE-M3 dense vector via explicit top-k thresholding because Ollama does not expose native sparse output.
3. ColBERT-style multi-vector — dense vector chunked into pseudo per-token embeddings for late interaction scoring.

Current Limitations

Ollama’s embedding API (/api/embed) currently returns only dense vectors. The sparse and multi-vec representations are derived from the dense output. They are not SPLADE outputs. When Ollama (or another backend) exposes native BGE-M3 sparse and per-token outputs, this implementation can use those representations directly without changing the retrieval interface.

The model bge-m3 produces 1024-dimensional dense embeddings.

Implementations

impl BgeM3Embedder

pub fn try_new(config: &EmbeddingConfig) -> Result<Self, MemoryError>

Create a new BgeM3Embedder from embedding config.

Uses bge-m3 as the model name with 1024 dimensions by default. The Ollama URL and timeout are taken from the config.

pub fn try_new_with_derive( config: &EmbeddingConfig, derive_config: BgeM3DeriveConfig, ) -> Result<Self, MemoryError>

Create a new BgeM3Embedder with custom derivation parameters.

pub fn with_params( base_url: &str, model: &str, dimensions: usize, batch_size: usize, timeout_secs: u64, derive_config: BgeM3DeriveConfig, ) -> Result<Self, MemoryError>

Create a BgeM3Embedder with explicit parameters (useful for testing without an EmbeddingConfig).

Trait Implementations

impl Embedder for BgeM3Embedder

BgeM3Embedder also implements the standard Embedder trait for compatibility, returning only the dense representation.

fn embed<'a>(&'a self, text: &'a str) -> EmbedFuture<'a>

Embed a single text. Returns a vector of f32.

fn embed_batch<'a>(&'a self, texts: Vec<String>) -> EmbedBatchFuture<'a>

Embed multiple texts in a batch.

fn model_name(&self) -> &str

The model name this embedder uses.

fn dimensions(&self) -> usize

Expected embedding dimensions.

fn embed_multi_optional<'a>( &'a self, text: &'a str, ) -> OptionalMultiEmbedFuture<'a>

Optionally produce dense and sparse representations in one model call.

fn embed_batch_multi_optional<'a>( &'a self, texts: Vec<String>, ) -> OptionalMultiEmbedBatchFuture<'a>

Optionally produce batched dense and sparse representations.

impl MultiFunctionEmbedder for BgeM3Embedder

fn embed_multi<'a>(&'a self, text: &'a str) -> MultiEmbedFuture<'a>

Embed a single text, returning dense, sparse, and multi-vec representations.

fn embed_batch_multi<'a>( &'a self, texts: Vec<String>, ) -> MultiEmbedBatchFuture<'a>

Embed multiple texts in a batch, returning all three representations per text.

fn model_name(&self) -> &str

The model name this embedder uses.

fn dimensions(&self) -> usize

Expected dense embedding dimensions.