elasticsearch - Is it possible to query Elastic Search with a feature vector? -

i'd store n-dimensional feature vector, e.g. <1.00, 0.34, 0.22, ..., 0>, each document, , provide feature vector query, results sorted in order of cosine similarity. possible elastic search?

i don't have answer particular elastic search because i've never used (i use lucene on elastic search built). however, i'm trying give generic answer question. there 2 standard ways obtain nearest vectors given query vector, described follows.

k-d tree

the first approach store vectors in memory of data structure supports nearest neighbour queries, e.g. k-d trees. k-d tree generalization of binary search tree in sense every level of binary search tree partitions 1 of k dimensions 2 parts. if have enough space load points in memory, possible apply nearest neighbour search algorithm on k-d trees obtain list of retrieved vectors sorted cosine similarity values. obvious disadvantage of method not scale huge sets of points, encountered in information retrieval.

inverted quantized vectors

the second approach use inverted quantized vectors. simple range-based quantization assigns pseudo-terms or labels real numbers of vector these can later indexed lucene (or matter elastic search).

for example, may assign label a range [0, 0.1), b range [0.1, 0.2) , on... sample vector in question encoded (j,d,c,..a). (because [.9,1] j, [0.3,0.4) d , on).

consequently, vector of real numbers transformed string (which can treated document) , hence indexed standard information retrieval (ir) tool. query vector transformed bag of pseudo-terms , 1 can compute set of other similar vectors in collection similar (in terms of cosine similarity or other measure) current one.

the main advantage of method scales massive collection of real numbered vectors. key disadvantage computed similarity values mere approximations true cosine similarities (due loss encountered in quantization). smaller quantization range achieves better performance @ cost of increased index size.