In the demo above, Smart Search turns short descriptions into useful actions. Type “video this week” to look for recent videos, describe an app's purpose instead of its name, or search for a page you visited in Chrome. Results are organized as Files, Apps, and Web, with a local result list available immediately.
That is the central problem: the user remembers intent, but the computer stores filenames, application names, timestamps, and URLs. A conventional text match helps when the words line up. A smarter search must bridge the gap without making every keystroke wait on a remote model.
What is Jev?
Jev is TypeSafe's first System One model. It is designed for decisions rather than prose generation: provide a state, ask bounded questions, and receive typed answers that application code can use directly.
TypeSafe exposes three question types: Choice selects among options you define, Noul estimates whether a statement is true, and Score evaluates an ordered rubric. Questions can be evaluated together in one call. Choice and Score include confidence information; Noul returns a probability. TypeSafe describes this as a fast, structured decision interface, but the model can still make a wrong judgment, so the application must decide how to use its output.
Jev does not crawl your Mac or find files by itself. Your code supplies candidate records. Jev judges the meaning of the query against those candidates; deterministic code still owns access, filtering, safety, and execution.
Why this is a good search problem for Jev
A launcher already has a finite set of actions: open this file, open that app, revisit a page, or search the web. The model need not invent an answer. It can choose which existing candidate best matches a phrase, estimate whether each candidate fits, and judge whether the user means one result or a set of results.
For example, “the PDF I just downloaded” requires more than matching “PDF”: “downloaded” points to when the file was added. “AI agent app” may describe an app's role rather than its title. Jev can compare these meanings across a shortlist, while normal code enforces the time window and ensures that only real, openable items can become actions. Multiple independent judgments fit naturally into TypeSafe's single structured request.
The implementation, from typing to opening
Gather locally. SuperNotch indexes installed apps, selected user folders, Chrome history and bookmarks, and Mac actions. Spotlight supplements file results, including indexed files on mounted volumes.
Filter before judging. A local fuzzy match returns results immediately. File types and phrases such as “this week” are interpreted in code, and an ordinary search sends only a bounded shortlist onward.
Ask Jev when enabled. After typing pauses, one request asks which candidate is the target, what action is intended, whether each result fits, and whether the query refers to one item or many. The response changes the ranking; it does not create new files or execute an action.
Keep control in the app. SuperNotch validates the response and uses confidence thresholds before promoting a result. The user chooses what to open. If the service is unavailable, rate-limited, or has no usable answer, local search remains available.
Responsiveness depends as much on engineering around the model as on the model itself. The app reuses its local index, debounces network judgments, cancels stale requests, caps candidate and query size, limits Spotlight work, and backs off after rate limits. Date calculations and file-type rules stay deterministic, so a model cannot silently redefine what “this week” means.
Privacy boundary
Local search works without a TypeSafe key. When you enable Jev ranking and provide a key, SuperNotch sends the query, bounded candidate metadata (such as names, details, type, and recency), and limited app context to TypeSafe. It does not send complete file bodies in that ranking request. Leave Jev ranking off if that metadata must stay on your Mac.
The product architecture above comes from the current SuperNotch implementation; Jev's model behavior is described in TypeSafe's documentation and its System One introduction.
Use the finished feature in SuperNotch
SuperNotch brings Smart Search into the same Mac notch toolkit as Clipboard History, screenshots, and screen recording. It is built for everyday use, with work already done on empty queries, ambiguous results, date buckets, external-volume results, Chrome source handling, stale requests, and responsive local fallback. These safeguards and performance limits make it practical to use directly; results still depend on what macOS and Chrome have actually indexed or recorded.
Find what you meant, faster.
Download SuperNotch to try Smart Search on your Mac. Local search works out of the box; Jev ranking is optional and needs your TypeSafe API key. A SuperNotch license is required to use the app.
Explore SuperNotch