Function: createColbertEmbedder()

createColbertEmbedder(config, runtime?): Promise<ColbertEmbedder>

Defined in: extensions/nlp/tasks/colbertEmbedding.ts:235

Creates a ColBERT (late-interaction) embedder, which maps a text to one vector per token rather than a single pooled vector. Relevance is then scored with ColbertEmbedder.maxSim instead of a cosine.

Inputs are built the way PyLate builds them: [BOS, prefix, ...tokens], with queries padded to a fixed length by expansion tokens and documents fed at their exact length. The per-token projection and L2 normalization are baked into the exported .pte.

Parameters

config

ColbertEmbedderModel

ColBERT embedder configuration containing the model and tokenizer paths and the tokenization settings. See ColbertEmbedderModel.

runtime?

any

Optional worklet runtime thread on which to run the model execution.

Returns

Promise<ColbertEmbedder>

A promise resolving to the instantiated ColbertEmbedder.

Throws

With code LOAD_FAILED if the model or tokenizer fails to load, SCHEMA_MISMATCH if the model schema does not match the ColBERT specification, or INVALID_ARGUMENT if a configured token is not in the vocabulary or queryLength does not fit the model.