Vettore.Encoding.Muvera (Vettore v0.3.2)

MUVERA-style fixed dimensional encodings for multi-vector retrieval.

This implements the public FDE boundary: query encodings sum vectors in each partition, document encodings average vectors in each partition, and both use the same deterministic partition configuration.

Summary

Types

config()

normalized_config()

vector()

Functions

encode_document(vectors, config \\ [])

Encodes document vectors by averaging projected vectors in each partition.

encode_query(vectors, config \\ [])

Encodes query vectors by summing projected vectors in each partition.

Types

config()

@type config() :: keyword()

normalized_config()

@type normalized_config() :: %{
  dimension: pos_integer(),
  num_repetitions: pos_integer(),
  num_simhash_projections: non_neg_integer(),
  seed: non_neg_integer(),
  projection_dimension: pos_integer(),
  final_projection_dimension: pos_integer() | nil
}

vector()

@type vector() :: [number()]

Functions

encode_document(vectors, config \\ [])

@spec encode_document([vector()], config()) :: {:ok, [float()]} | {:error, term()}

Encodes document vectors by averaging projected vectors in each partition.

encode_query(vectors, config \\ [])

@spec encode_query([vector()], config()) :: {:ok, [float()]} | {:error, term()}

Encodes query vectors by summing projected vectors in each partition.