Browser-Based Image Search Using EmbeddingGemma 2 and WebGPU
A developer ported Google's EmbeddingGemma 2 vision and text towers to ruNNtime, a WebGPU inference engine written in TypeScript. This implementation allows users to search photo galleries using natural language descriptions directly inside their web browser. This project demonstrates fully client-side multimodal AI inference, enabling private and fast image retrieval without sending media or queries to external servers. It showcases the expanding capability of WebGPU to run advanced computer vision and embedding models locally on edge devices. The system maps both images and text queries into a shared 768-dimensional vector space to compute similarity directly on the user's GPU. Built using Software Mansion's ruNNtime framework, the setup leverages WebGPU for hardware acceleration within modern web browsers.
## BACKGROUND
EmbeddingGemma is an open embedding model developed by Google DeepMind designed to convert text and visual data into dense vector representations for retrieval tasks. WebGPU is a W3C web standard that gives web applications low-level access to modern GPU hardware, enabling high-performance machine learning inference directly inside the browser.