US 12,423,310 B1
Filter direction modifications to vector-based search
Milind Arjanbhai Shyani, Newark, CA (US); Yonatan Naamad, Sunnyvale, CA (US); and Supriya Nagesh, Sunnyvale, CA (US)
Assigned to Amazon Technologies, Inc., Seattle, WA (US)
Filed by Amazon Technologies, Inc., Seattle, WA (US)
Filed on Dec. 15, 2023, as Appl. No. 18/542,378.
Int. Cl. G06F 16/2457 (2019.01); G06F 16/22 (2019.01); G06F 16/248 (2019.01)
CPC G06F 16/24573 (2019.01) [G06F 16/2237 (2019.01); G06F 16/248 (2019.01)] 20 Claims
OG exemplary drawing
 
5. A method, comprising:
obtaining a query data item and a data item filter to perform a data item search on plurality of data items;
generating a query vector that represents the query data item in feature space using an encoder of a trained machine learning model that encodes data items in the feature space;
obtaining a filter vector that represents the data item filter in the feature space from a matrix of filter vectors generated for a plurality of data item filters, wherein the matrix of filter vectors provides respective directions corresponding to the plurality of data items in the feature space;
generating a modified query vector based on the filter vector and the query vector;
performing a nearest neighbor search technique in the feature space using the modified query vector to identify one or more data items of the plurality of data items in a data item index comprising respective index vectors for individual ones of the plurality of data items; and
providing the one or more data items as result for the data item search.