Struct Bgem3Provider 

pub struct Bgem3Provider { /* private fields */ }

BGE-M3 joint embedder — dense + sparse from one pass.

Memory

fastembed’s BGE-M3 graph always emits a third, ColBERT (multi-vector) output, and accumulates it for every text passed in a single call — roughly seq_len × 1024 × 4 bytes each, about 2 MB for a 512-token chunk. Handing it a million chunks at once would therefore exhaust memory long before it finished. This provider slices its input into with_batch_cap-sized runs and drops the ColBERT tensors as soon as each run is converted, so bulk indexing stays flat.

Model

fastembed ships BGE-M3 as an int8-quantized ONNX export (gpahal/bge-m3-onnx-int8), not full precision.

Implementations

impl Bgem3Provider

pub fn new(cache_dir: Option<PathBuf>) -> Result<Self>

Load BGE-M3, downloading it on first use.

cache_dir overrides the HuggingFace model cache directory.

pub fn with_max_length( max_length: usize, cache_dir: Option<PathBuf>, ) -> Result<Self>

Load BGE-M3 with an explicit maximum sequence length. Longer inputs are truncated; BGE-M3 itself supports up to 8192 tokens.

pub fn with_batch_cap(self, cap: usize) -> Self

Cap how many texts are sent to the model per call.

This bounds peak memory: the discarded ColBERT output is accumulated per call, so a larger cap trades memory for slightly fewer invocations. A cap of 0 is treated as 1.

pub fn with_sparse_top_k(self, k: usize) -> Self

Keep only the k highest-magnitude weights of each sparse vector.

Set this when the target store bounds non-zero elements — pgvector refuses to build an HNSW index over a sparsevec beyond its cap. Off by default, since the limit belongs to the backend, not the model.

pub fn dense_dim(&self) -> usize

Dense vector width.

pub fn vocab_size(&self) -> usize

Vocabulary size — the sparsevec(N) dimension for the sparse side.

pub fn embed(&self, texts: &[&str]) -> Result<Vec<JointEmbedding>>

Embed texts, returning one dense + sparse pair each, in input order.