Watch: Inside Chrome’s Semantic Engine: A Technical Analysis of History Embeddings

Google Chrome has introduced a powerful new way to search your browsing history using natural language. Instead of trying to remember exact keywords, you can ask Chrome simple questions, like, "What was that ice cream shop I looked at last week?"

To make this work, Chrome relies on a sophisticated, local processing pipeline. First, an internal tool breaks down web pages into semantically meaningful text passages. It carefully analyzes the structure of each webpage, grouping related text together while ignoring short, irrelevant fragments.

Next, Chrome converts these passages into high-dimensional mathematical vectors, or embeddings, which capture the actual meaning of the words. Each page is turned into a set of fifteen-hundred-dimensional vectors.

For privacy and speed, all of this happens directly on your device. The data is compressed, encrypted, and stored locally in Chrome's history database. Private browsing is completely ignored, and no browsing data is sent back to Google.

When you search, Chrome compares your query to these stored vectors. An intelligent classifier figures out your intent, and a helper system can even generate direct answers using your history as a personal knowledge base. To keep your computer running smoothly, Chrome schedules this heavy lifting during idle moments, using hardware optimizations to keep your browsing seamless. It is a highly secure, efficient system that turns your history into a smart, private search engine.