Does the assistant send my pages anywhere?
No. It runs in your browser. The only network use is downloading the model once, from its public host, after you turn it on.
Ask your pages. Nothing leaves the device.
Vellum’s assistant is off until you turn it on. Then a language model downloads once and runs in your browser. It answers from what you wrote, shows where it found it, and can run code for you, with your approval.
On your device, after you opt in
Nothing AI-related downloads until you turn the assistant on. Then its model is fetched once from its public host and cached by the browser; it runs on the GPU with WebGPU, or as a smaller model on the CPU. Your questions, the answers and the search index it builds stay in the browser’s storage on this device.
The model is a small open-weights model running locally: it is slower and less capable than large cloud models, and it can be wrong. Its citations show you where to check.
Answers with sources
Every answer cites the pages it used as numbered chips; a chip opens the page. Ask across all pages, one notebook or only the open page. Files count too: text inside PDFs, Word documents and slides is indexed page by page, so a citation can point at the exact page of a PDF.
More than chat
- Summaries and outlines of the open page, or a continuation in your words.
- Study guides and audio overviews of a notebook.
- Explain and Fix for code cells: a fix arrives as a diff you accept or reject.
- Run code: to compute an answer, the assistant can write code and run it, after you approve each run on a card in the answer. The transcript shows the code and its output.
In the app
- Ask your pages Answers from your own pages, with sources. Runs on this device.
- Ask about one page Ask about this page, a notebook, or everything.
- Summaries and outlines Summarize, outline or continue the open page.
- Study guides and audio Turn a notebook into a study guide or listening overview.
- Explain and Fix The on-device assistant explains code and suggests fixes.
- The assistant runs code It can compute answers, after you approve each run.
Questions
What does it need?
A current browser. It is fastest with WebGPU (recent Chrome and Edge); without it, a smaller model runs on the CPU.
Can it change my pages?
Only when you ask for a summary, outline or continuation and keep it. Code it wants to run waits for your approval.