Chrome History Embeddings
Chrome's local system that turns pages you visit into vectors, enabling natural-language search of your own browsing history on-device.
Have you ever tried to find a website you visited last week, but you couldn't remember the name? Chrome is solving this with a new AI-powered history search that lets you find past pages using natural-language queries. You can simply search for something like "that ice cream shop I looked at last week," and Chrome will find it.
Under the hood, Chrome does this by turning the pages you visit into high-dimensional vectors and storing them right on your device. It all starts with a document chunker that scans the page. It walks through the website's structure to group the content into meaningful passages.
Next, Chrome converts these text chunks into fifteen-hundred-and-forty-dimensional vectors. When you search, the browser runs a local vector search, scoring relevance by calculating the dot product of normalized vectors, a method known as cosine similarity.
This on-device process offers a rare look at how Google handles embedding and chunking in the real world. It is the very same family of techniques Google uses for its main search engine, now running privately and directly on your own computer.
Chrome history embeddings are the vectors behind Chrome's AI-powered history search, which lets you find past pages with natural-language queries like "what was that ice cream shop I looked at last week?" It processes pages you visit into semantic passages, converts them to high-dimensional vectors, and stores them locally.
Decoded from Chromium's source, the pipeline centres on a DocumentChunker that walks the DOM with a recursive tree-walking algorithm, respecting semantic structure to aggregate content into meaningful passages before embedding them into 1540-dimensional vectors. Search then runs on-device via vector search.
It's a rare, concrete look at Google's real embedding and chunking choices — the same family of techniques used across search. Relevance is scored with a dot product over normalised vectors (see cosine similarity), and it's one of the many on-device models Chrome now runs.