Type Alias: ColbertEmbedder

ColbertEmbedder = object

Defined in: extensions/nlp/tasks/colbertEmbedding.ts:100

ColBERT embedding task runner.

Properties

dispose()

readonly dispose: () => void

Defined in: extensions/nlp/tasks/colbertEmbedding.ts:104

Releases all allocated native resources.

Returns

void

---

embed()

readonly embed: (input, kind) => Promise<ColbertTokenEmbeddings>

Defined in: extensions/nlp/tasks/colbertEmbedding.ts:120

Asynchronously computes the per-token embeddings of the given text.

A query yields exactly ColbertEmbedderOptions.queryLength vectors. A document yields one vector per token, minus the skiplisted ones, and is truncated to the model's maximum sequence length. Score a query against a document with ColbertEmbedder.maxSim.

Parameters

input

string

The text to embed.

kind

ColbertInputKind

Whether input is a search query or a document to search.

Returns

Promise<ColbertTokenEmbeddings>

A promise resolving to the token embeddings.

Throws

With code INVALID_ARGUMENT if the text tokenizes to zero tokens or kind is unknown, RESOURCE_BUSY if the model is in use, or RESOURCE_DISPOSED if disposed.

---

embedWorklet()

readonly embedWorklet: (input, kind) => ColbertTokenEmbeddings

Defined in: extensions/nlp/tasks/colbertEmbedding.ts:126

Synchronous version of embed to be executed directly on the caller or worklet thread.

Parameters

input

string

kind

ColbertInputKind

Returns

ColbertTokenEmbeddings

---

maxSim()

readonly maxSim: (query, document) => number

Defined in: extensions/nlp/tasks/colbertEmbedding.ts:141

Computes the late-interaction (MaxSim) relevance of a document to a query: for every query vector, the highest dot product with any document vector, summed over the query. With L2-normalized vectors each term is a cosine similarity, so the score lies in [-numTokens, numTokens] of the query. It does not run the model, so it also works on stored embeddings.

Parameters

query

ColbertTokenEmbeddings

The query token embeddings.

document

ColbertTokenEmbeddings

The document token embeddings.

Returns

number

The MaxSim score; higher is more relevant. A document with no vectors scores 0.

Throws

With code INVALID_ARGUMENT if the two embeddings have different vector dimensions.