VectorDistance (NoSQL query)

  • Article

APPLIES TO: NoSQL

Returns the similarity score between two specified vectors.

Syntax

VectorDistance(<vector_expr_1>, <vector_expr_2>, <bool_expr>, <obj_expr>)

Arguments

Parameter Description
vector_expr_1 An array of float32 or smaller.
vector_expr_2 An array of float32 or smaller.
bool_expr An optional boolean specifying how the computed value is used in an ORDER BY expression. If true, then brute force is used. A value of false uses any index defined on the vector property, if it exists. Default value is false.
obj_expr An optional JSON formatted object literal used to specify options for the vector distance calculation. Valid items include distanceFunction and dataType, and searchListSizeMultiplier.

Supported parameters for the optional obj_expr

Parameter Description
distanceFunction The metric used to compute distance/similarity.
dataType The data type of the vectors. float32, int8, uint8 values. Default value is float32.
searchListSizeMultiplier An integer specifying the size of the search list when conducting a vector search. Increasing this may improve accuracy at the expense of RU cost and latency. Min=1, Default=5, Max=100.

Supported metrics for distanceFunction are:

  • cosine, which has values from -1 (least similar) to +1 (most similar).
  • dotproduct, which has values from -∞ (-inf) (least similar) to +∞ (+inf) (most similar).
  • euclidean, which has values from 0 (most similar) to +∞ (+inf) (least similar).

Return types

Returns a numeric expression that enumerates the similarity score between two expressions.

Examples

This first example shows a top 10 vector search query with only the required arguments. One property is projected, along with the score returned by VectorDistance. Then, we user an ORDER BY clause to sort VectorDistance scores in order from most similar to least.

SELECT TOP 10 s.name, VectorDistance(c.vector1, <query_vector>)
FROM c
ORDER BY VectorDistance(c.vector1, <query_vector>)

This next example also includes optional arguments for VectorDistance

SELECT TOP 10 s.name, VectorDistance(c.vector1, <query_vector>, true, {'distanceFunction':'cosine', 'dataType':'float32',})
FROM c
ORDER BY VectorDistance(c.vector1, <query_vector>, true, {'distanceFunction':'cosine', 'dataType':'float32',})

Remarks

  • This function requires enrollment in the Azure Cosmos DB NoSQL Vector Search preview feature.
  • This function benefits from a vector index
  • if false is given as the optional bool_expr, then the vector index defined on the path is used, if one exists. If no index is defined on the vector path, then this function reverts to full scan and incurs higher RU charges and higher latency than if using a vector index.
  • When VectorDistance is used in an ORDER BY clause, no direction needs to be specified for the ORDER BY as the results are always sorted in order of most similar (first) to least similar (last) based on the similarity metric used.
  • The result is expressed as a similarity score.

Related content