The algorithm is based on the assumption that two real numbers can be considered equal within certain equality distance. Then quantization is used for comparison. To make sure points near or at quantization borders are also comparable, a vector can be discretized into more than one hash, as described here (also as PDF). The method indirectly clusters given vectors by hypercubes and their neighbourhoods. It is exhaustive within a set precision.

The algorithm assumes a uniform and normalized vector space - without complex manifolds or very diverse properties of dimensions, which can potentially complicate search. But even with these complications, sufficiently large hupercubes (small number of buckets) will probably work fine for prefiltering or sequencial filtering by smaller-dimentional sub-spaces, as briefly mentioned in the article.

It has not been tested on very high-dimensional vectors - but they may produce impractically large hash sets. The linked example below uses only 9 dimensions.

How to use

Normalize each component of your input vectors to the same min/max value range. Use these min/max values in the parameters settings.
Provided a float vector []float64, use CentralCube and CubeSet functions to generate hypercube coordinates []int and [][]int.
Generate a DecimalHash/FNV1aHash and HashSet for corresponding central hash and hash set from the hypercube coordinates above. The difference between one hash and a hash set is that one corresponds to a hash-table record and the other to a query or vice versa, depending on performance/memory preference. There are 2 alternative hash functions: DecimalHash and FNV1aHash. DecimalHash does not have collisions, but is not suitable for cases with large number of buckets or dimensions. FNV1aHash is applicable for all cases. Hash collisions can be progressively eliminated by using custom hash functions or verifying similarity with the Euclidean metric.

Example for similar image search and clustering.

Go doc for full code documentation.