Source: Qdrant docs: Vectors (multivectors)

Qdrant's vector concepts page: native multivector points scored with MaxSim, configuration, and the hard per-point size limit.

Contents

Summary

Documents how a Qdrant point can hold a matrix of same-sized vectors instead of one vector, used either for several views of one object or for late-interaction token embeddings such as ColBERT.

Key claims

  • Configure a dense vector with multivector_config and comparator: max_sim; MaxSim is currently the only comparator.
  • MaxSim sums, over each query vector, its best similarity to any document vector, and returns one score per point.
  • Hard limit: number of sub-vectors x vector size must stay under 1,048,576 (for 4096-dim vectors, at most 255 sub-vectors). Workarounds: pool token vectors, or split a document into several points and aggregate.
  • Compressed datatypes (turbo16, turbo8, turbo4) support multivectors.

Feeds

Late-interaction retrieval

Suggest changes →