Package io.nosqlbench.virtdata.lib.vectors.dnn

package io.nosqlbench.virtdata.lib.vectors.dnn

This package contains experimental support for new methods for testing vector stores. The primary method employed is functional mapping of ordinal spaces to vector spaces. In this way, closed-form functions can be used to synthesize vectors and provably correct neighborhoods as if they were defined in a static dataset. This allows for arbitrary testing scenarios to be created and used immediately and with no need to regenerate or compute any data beforehand.

The original concept for this was derived by Shaunak Das, in the form of (Das) Direct Nearest Neighbor. Additional methods have been implemented using this technique to include additional space mappings for other vector distance functions.

The testing methods enabled by this approach include:

  1. Generation of a population of vectors which are enumerable and stable with respect to their ordinal addresses.
  2. Generation of ordered subsets of this population which maintain a unique local ordering in terms of the selected distance function, otherwise known as rank for KNN queries.
  3. Validation of results for nearest neighborhood queries, using synthetic results computed on the fly as the basis for correctness.

The vector spaces constructed in this way are not intended nor guaranteed to be dimensionally disperse. They are meant to provide an algebraic basis for exercising vector storage systems with increasing cardinality of vectors. This means that vector stores can be tested to incrementally higher limits while their performance and accuracy are both measured.

Each vector scheme in this method has the following properties:

  • Each virtual vector space is defined by a set of parameters which are used as inputs to the mapping functions. The space, and the definition of valid vectors in a neighborhood depend on these for stability and correctness. Thus each space is explicitly defined by and inseparable from its parameters.
  • All vectors within the space are enumerable. Each increasing ordinal value describes a new and distinct vector. The value of this vector is deterministic within the parameters of the space.
  • Each vector within a space is a valid query vector which implies a correct set of distance-ranked neighbors up to some neighborhood size for the related distance function.
  • Nearest neighbors may have equal distance in some cases, for which ties are accommodated in testing assertions. Suppose the distance from v10 to v5 is the same as the distance from v10 to v15, then both v5 and v15 should be interchangeable as correct elements in any KNN results for query vector v10, provided that their distances are within the top K results as otherwise expected.

TBD: Explain the above in terms of specific implementations and parameters.