generating a similarity graph based at least in part on distances between vector representations of a first plurality of content items in a vector space;

determining, based at least in part on the similarity graph, a first content item of the first plurality of content items as a representative item for a cluster of content items;

receiving a user query;

generating, based at least in part on the user query, a query vector that is a first vector representation of the user query;

comparing the query vector with an item vector, wherein the item vector comprises a second vector representation of the first content item;

determining, based at least in part on the comparing the query vector with the item vector, the cluster of content items as a candidate cluster;

generating a sparse representation of the query vector, wherein the sparse representation comprises a dimension with a null or non-null value, wherein the generating expands the query vector and inserts at least one of a null or zero value for numerical values in the query vector representing a feature without equivalent in the item vector;

partitioning the cluster of content items based on non-null value associations with the dimension, wherein each of the content items in a first partition of the cluster of content items comprises a respective vector representation having a respective non-null value for a dimension;

determining a similarity value between the query vector and the respective vector representation of at least one of the content items in the first partition of the cluster of content items;

determining based upon the similarity value, a candidate item in the first partition of the cluster of content items; and

sending, to a computing device based at least in part on the user query, search results comprising the candidate item.