Described herein is a technology that facilitates efficient large-scale similarity-based retrieval. In several embodiments documents, images, and/or other multimedia files are compactly represented and efficiently indexed to enable robust search using a long-query in a large-scale corpus. As described herein, these techniques include performing decomposition of a file, e.g., an image, a document containing an image, or a document-like representation of an image. The techniques use dimension reduction to obtain three parts, low-dimensional representations (major semantics), file specific terms (minor semantics), and background words, representing the major semantics in a feature vector and the minor semantics as keywords. Using the techniques described, file vectors are matched in a topic model and the results ranked based on the keywords.