identifying an original query comprising an image description;

tokenizing the original query to obtain a plurality of original tokens;

generating a plurality of expanded queries by generating a plurality of additional phrases based on the original query using a causal language model (CLM), wherein each of the plurality of expanded queries comprises an expanded image description that includes the original query and an additional phrase from the plurality of additional phrases, wherein the CLM generates each of the plurality of additional phrases by generating a plurality of sequences of tokens, respectively, and wherein each token of each sequence of the plurality of sequences of tokens is generated based on the plurality of original tokens and a sequence of previously generated tokens;

inserting a mask token at a target location of each of the plurality of expanded queries;

replacing the mask token with an insertion phrase using a masked language model (MLM) to obtain a plurality of modified queries, respectively; and

providing a plurality of images in response to the original query, wherein the plurality of images are associated with the plurality of modified queries, respectively.