Module hybrid 

Single-pass hybrid embedding: the HybridEmbeddingModel trait and its HeadSet / HybridEmbedResult types.

Some exports — notably BGE-M3 (aapot/bge-m3-onnx) — fuse a dense head, a learned-sparse head, and a multi-vector (ColBERT) head into one graph. The per-task EmbeddingModel, SparseEmbeddingModel, and MultiVectorEmbeddingModel handles each load their own session and run their own forward pass, so serving all three means loading the weights three times and running the graph three times.

This trait collapses that to one: a single shared encoder forward pass whose outputs are post-processed into every requested head. It exists only for graphs that actually expose multiple heads (declared by a hybrid preset); single-head models keep using the per-task handles. The caller picks which heads to materialize with a HeadSet, mirroring how NlpTasks selects NLP heads on a single pass.

Structs

HeadSet

Selects which embedding heads a HybridEmbeddingModel should populate.

HybridEmbedResult

Dense, sparse, and multi-vector embeddings from a single forward pass.

Traits

HybridEmbeddingModel

A model that produces several embedding heads from one shared forward pass.