How image search works at Dropbox

This post details the technical approach to implementing image content search at Dropbox. It explains the core idea of using a relevance function based on image classification and word vectors. The post describes how image classifiers (EfficientNet trained on OpenImages) generate category scores, and how word vectors (ConceptNet Numberbatch) are used to represent search queries in the same category space. It outlines the mathematical formulation for projecting query word vectors into category space and calculating relevance scores. The production architecture leverages the Nautilus search engine, explaining how inverted indexes are used to efficiently retrieve relevant images based on image content metadata.