- ON clause
- Accepts the ReferenceTable clause.
- RefIDColumn
- Specify the name of the reference table column that contains identifiers of the reference table vectors.
- RefFeatureColumns
- List of reference table column names that contain features of the reference table vectors.You can specify up to 2018 feature columns.
When providing embedding types, TargetFeatureColumns and RefFeatureColumns must have 1 column.
- DistanceMeasure
- Specifies the distance type to compute between the target and the reference vector:
- Cosine: Cosine distance between the target vector and the reference vector.
- Euclidean: Euclidean distance between the target vector and the reference vector.
- Manhattan: Manhattan distance between the target vector and the reference vector.
- Dot product: Dot product distance between the target vector and the reference vector.Both target and ref vectors must be normalized for dot product distance.
- Minkowski: Minkowski distance between the target vector and the reference vector.
Default: cosine
- TopK
- Maximum number of closest reference vectors to include in the output table for each target vector. The value k is an integer between 1 and 1024.
Default value: 1024
Important: TD_VectorDistance returns N2 output if you use the TopK value as -1 because the function includes all reference vectors in the output table.TopK and SearchThresholds cannot be provided together. If both are omitted, then topk = 1024 will be considered. - PValue
- Specifies the integer value needed to compute the minkowski distance.
Default value = 2
Range 1-INT_MAX
- SearchThresholds
- Specifies the search threshold between a pair of target and reference vectors.
The pair does not appear in the output table if the distance exceeds the search threshold.
The pair does not appear in the output table if the similarity is less than the search threshold.
If DistanceMeasure specifies multiple measures, then SearchThresholds must specify a threshold for each measure.
The ith threshold corresponds to the ith measure.
SearchThresholds and TopK cannot be provided together, otherwise an error is reported. If both are omitted, then topk = 1024 will be considered. - LargeReferenceInput
- If the reference table is much larger than the target table, then passing true in this syntax element will treat the reference table as 'Partition By any' instead of 'Dimension', and the targetTable will be treated as Dimension. You must pass the reference table as 'Partition By any' and target table as Dimension.
True (true, t, yes, y, 1) False (false, f, no, n, 0)
Default: False (the reference table will be Dimension only)
- OutputSimilarity
- Specifies true if you want similarity instead of distance.
True (true, t, yes, y, 1) False (false, f, no, n, 0)
Default: False (the function will emit the distance)
Similarity is computed in the following manner:- Cosine_similarity = 1 - Cosine_distance
- Euclidean_similarity = 1/(1 + Euclidean_distance)
- Manhattan_similarity = 1/(1 + Manhattan_distance)
- Dot product_similarity = 1 - DotProduct_distance Both target and ref vectors must be normalized for dot product similarity.
- Minkowski_similarity = 1/(1 + minkowski_distance)
- EmbeddingSize
- Specify the embedding size of the vectors.
Max value = 4096
- UseSIMD
- Specifies whether to invoke SIMD instruction set for computing vector operations (distance between vectors).
Default: false