What does “AI on your device” actually mean?
A plain-language look at model downloads, local inference, and what leaves your browser.
The model comes to you
When you choose to enable local AI, your browser downloads model files from the model provider. These requests expose ordinary network information such as your IP address to that provider. Your sleep records and questions are not included in those requests.
The conversation stays here
KnowNaps can run Qwen3 1.7B through WebLLM in a browser worker using WebGPU. Common questions are calculated immediately. For unfamiliar wording, optional AI creates a constrained query over your complete imported history. The app then selects the matching records and calculates the answer, with dates and charts. There is no cloud chat endpoint or account connection.
Small models have limits
Language models can misunderstand a question. KnowNaps validates their query parameters and calculates measurements in code rather than displaying invented numerical prose. Every answer shows its scope; explicit dates are protected from model changes. Unsupported questions ask for clarification. The optional model needs about a 1 GB first download and about 2.1 GB GPU memory. On unsupported devices, direct questions and all charts still work.
You control storage
Imported records stay in this tab by default. Remember on this device is optional and stores records in this browser, unencrypted. Reloading or closing the tab clears session-only records; use Clear my data to remove saved workspace records too. Model download caches are separate from personal records.
General education only. Persistent sleep problems deserve a conversation with your healthcare professional.