Sparse embeddings

Sparse models encode terms into parallel index/value arrays. Use them for lexical signals or combine them with dense semantic vectors.

Index passages

const sparse = await createFastEmbedSparseEmbeddingModel()

const passages = await sparse.embedTexts([
  'Password reset links expire after thirty minutes.',
  'Enterprise accounts include priority support.',
])

embedTexts() calls FastEmbed’s passage encoder and returns { document, vector } values.

Encode a query

const query = await sparse.embedQuery('reset link expiration')

Query encoding uses FastEmbed’s distinct query path. Do not index documents with embedQuery() or search with a passage vector unless the model documentation explicitly calls for it.

Hybrid retrieval

A typical hybrid collection stores:

• one named dense vector from FastEmbedEmbeddingModel;
• one named sparse vector from FastEmbedSparseEmbeddingModel;
• source text and metadata for filtering/citations.

The vector store owns fusion and ranking. The FastEmbed adapter only produces vectors. For Qdrant, configure a hybrid collection and an RRF search path through @anvia/qdrant.

Validation and indexing

The adapter checks that sparse vectors contain numeric indices and values with matching lengths. It preserves provider order and values; it does not deduplicate, sort, or renormalize them.

Record the sparse model name with the collection. Changing the model requires re-embedding indexed passages and query configuration together.