Package version:

Interface BinaryQuantizationCompression

Contains configuration options specific to the binary quantization compression method used during indexing and querying.

interface BinaryQuantizationCompression {
    compressionName: string;
    defaultOversampling?: number;
    kind: "binaryQuantization";
    rerankWithOriginalVectors?: boolean;
    rescoringOptions?: RescoringOptions;
    truncationDimension?: number;
}

Hierarchy (view full)

Properties

compressionName defaultOversampling? kind rerankWithOriginalVectors? rescoringOptions? truncationDimension?

Properties

compressionName

compressionName: string

The name to associate with this particular configuration.

OptionaldefaultOversampling

defaultOversampling?: number

Default oversampling factor. Oversampling will internally request more documents (specified by this multiplier) in the initial search. This increases the set of results that will be reranked using recomputed similarity scores from full-precision vectors. Minimum value is 1, meaning no oversampling (1x). This parameter can only be set when rerankWithOriginalVectors is true. Higher values improve recall at the expense of latency.

kind

kind

Polymorphic discriminator, which specifies the different types this object can be

OptionalrerankWithOriginalVectors

rerankWithOriginalVectors?: boolean

If set to true, once the ordered set of results calculated using compressed vectors are obtained, they will be reranked again by recalculating the full-precision similarity scores. This will improve recall at the expense of latency.

OptionalrescoringOptions

rescoringOptions?: RescoringOptions

Contains the options for rescoring.

OptionaltruncationDimension

truncationDimension?: number

The number of dimensions to truncate the vectors to. Truncating the vectors reduces the size of the vectors and the amount of data that needs to be transferred during search. This can save storage cost and improve search performance at the expense of recall. It should be only used for embeddings trained with Matryoshka Representation Learning (MRL) such as OpenAI text-embedding-3-large (small). The default value is null, which means no truncation.

Settings

Member Visibility