Configuration

Dense and sparse factories accept parallel option shapes.

Dense model

const dense = await createFastEmbedEmbeddingModel({
  model: 'fast-bge-base-en-v1.5',
  maxBatchSize: 64,
  initOptions: {
    cacheDir: './.cache/fastembed',
    maxLength: 512,
    showDownloadProgress: false,
    executionProviders: ['cpu'],
  },
})

Sparse model

const sparse = await createFastEmbedSparseEmbeddingModel({
  model: 'prithivida/Splade_PP_en_v1',
  maxBatchSize: 64,
  initOptions: {
    cacheDir: './.cache/fastembed',
    showDownloadProgress: false,
  },
})

executionProviders uses the ExecutionProvider type from fastembed; valid values depend on the installed runtime. Do not copy an example provider into production without confirming support in the target image.

Option behavior

• model selects the FastEmbed enum-backed model.
• maxBatchSize controls adapter batching and is clamped to an integer of at least one.
• cacheDir controls model asset placement.
• maxLength and other initialization options are passed to FastEmbed.
• initOptions are forwarded and the selected top-level model is also added as model; avoid supplying a conflicting modelName.

Direct construction

Use new FastEmbedEmbeddingModel(runtime, options) or the sparse equivalent when the application already owns an initialized runtime or tests need a fake:

const model = new FastEmbedEmbeddingModel(runtime, {
  model: 'fast-bge-small-en-v1.5',
})

Direct construction does not download or initialize model assets. The supplied runtime must honor the documented async iterable shape.