Mist is the module that surprised us most in dogfooding. We built it expecting users to lean on web search; they lean on local first. Every day.
The three corpora
Mist searches across three corpora, in this priority order:
- Local notes. Anything you've typed into Nimbus8 or any chat, captured from Stratus, or attached from the Files app and told the system to index.
- On-device files. PDFs, images (via OCR), and text in any folder you've granted permission to index.
- Optional web. Off by default. Toggle in the composer. When on, it routes through a privacy-preserving search partner that doesn't receive your device identifier.
The answer you see is fused from all three, but each sentence carries a citation pill indicating which corpus it came from. You never lose track of whether the thing you're reading is from your own archive or from the open internet.
Why local wins so often
A surprising share of the questions people ask a search tool are answerable from things they've already seen. "What was that paper about compositional perception we were reading?" "Which receipt has the delivery date on it?" "Where did I write that note about the Q2 roadmap?" For all of these, the web is the wrong place to look; the answer is in your own stuff, and a search tool that can't look there is fundamentally limited.
Mist's local ranker uses a small on-device embedding model (gte-small, quantized). Indexing happens incrementally as you add content; it's fast enough to keep up with typing and doesn't require a background task.
How citations are rendered
Each citation is a pill attached to the sentence it supports. Tapping the pill expands a card showing the source: the original note with the matching passage highlighted, or the file thumbnail, or the web page's title + URL + favicon. You can tap through to the source in one hop. Citations are ordered by the ranker's confidence; the top one is the one most likely to be the actual answer.
We don't try to hide the provenance the way some AI search products do. If the answer is an amalgam of three sources, the three citations are there in line. If the answer is thin ("the model guessed"), the citation set is thin too, and the answer is rendered with a lower-confidence visual treatment.
The opt-in web path
When you flip the web toggle, Mist queries our search partner with the parsed query only. No device ID, no user agent beyond "Nimbus8/<version>", no account. Results come back, get embedded on-device, and get re-ranked alongside the local corpus. The web results are visually distinguished (a different pill color, a small globe icon) so you always know which sentences came from outside your device.
What Mist doesn't do
Mist doesn't scrape your email, your messages, or any app's data we don't have explicit permission to index. Mist doesn't cache web pages server-side. Mist doesn't build a searchable profile of you on anyone else's machine. Mist's index file is in your app sandbox; deleting Nimbus8 deletes the index with it.
Tagged Mist · Published Feb 28, 2026 · Back to all posts