Configuration

const embeddings = await loadTransformersEmbeddingModel({
  modelId: 'Xenova/all-MiniLM-L6-v2',
  pooling: 'mean',
  normalize: true,
  maxBatchSize: 16,
})

Options

Option	Default	Effect
modelId	Required; DEFAULT_TRANSFORMERS_EMBEDDING_MODEL exports Xenova/all-MiniLM-L6-v2 as the recommended default	Model passed to the feature-extraction pipeline. Omitting it or passing an empty value throws a TypeError.
pooling	'mean'	Chooses mean or CLS output pooling.
normalize	true	Requests normalized vectors from the pipeline.
maxBatchSize	16	Exposed batch-size metadata. Values below 1 (or non-integers) throw TypeError: maxBatchSize must be a positive safe integer; the adapter does not clamp.

The current embedTexts() implementation calls the extractor once with the full texts array. maxBatchSize is part of the Anvia model contract but does not split that call inside this adapter. If strict batching is required, split input in application code or inject a pipeline that owns batching.

Inject a pipeline

const embeddings = adaptTransformersEmbeddingModel({
  runtime: extractor,
  modelId: 'company/model',
  pooling: 'cls',
  normalize: true,
})

The extractor must accept (texts, { pooling, normalize }) and resolve an object whose tolist() returns the vectors. Adapting it does not load a model.

Index compatibility

Pooling and normalization are part of vector semantics. Store them with the model ID in collection metadata. Do not query an index created with mean-normalized output using CLS or unnormalized output.