@anvia/transformers API reference

Import every public symbol from @anvia/transformers. The package has no public subpath exports.

Types

type TransformersPooling = 'mean' | 'cls'

type TransformersFeatureExtractionPipeline = (
  texts: string[],
  options: {
    pooling: TransformersPooling
    normalize: boolean
  },
) => Promise<{
  tolist(): unknown
}>

type TransformersEmbeddingModelOptions = {
  model?: string
  pooling?: TransformersPooling
  normalize?: boolean
  maxBatchSize?: number
}

TransformersFeatureExtractionPipeline is the minimum public contract accepted by direct model construction. It allows an official Transformers.js pipeline, wrapper, or test double to be injected.

Constant

const DEFAULT_TRANSFORMERS_EMBEDDING_MODEL = 'Xenova/all-MiniLM-L6-v2'

TransformersEmbeddingModel

class TransformersEmbeddingModel implements EmbeddingModel {
  readonly model: string
  readonly maxBatchSize: number

  constructor(
    extractor: TransformersFeatureExtractionPipeline,
    options?: TransformersEmbeddingModelOptions,
  )

  static create(
    options?: TransformersEmbeddingModelOptions,
  ): Promise<TransformersEmbeddingModel>

  embedTexts(texts: string[]): Promise<Embedding[]>
}

create() loads a Transformers.js feature-extraction pipeline using options.model or the default. Pooling defaults to mean, normalization defaults to true, and the adapter batches according to maxBatchSize.

Direct construction uses the supplied extractor without loading a model. embedTexts() converts tolist() output into Anvia embeddings and rejects malformed vectors or a vector count that differs from the input count.

Factory

function createTransformersEmbeddingModel(
  options?: TransformersEmbeddingModelOptions,
): Promise<TransformersEmbeddingModel>

The factory is a convenience wrapper around TransformersEmbeddingModel.create(options) and returns a fully initialized model.