Search searches the text of every page in your space. Open it from the command palette or press Ctrl-Shift-f (Cmd-Shift-f on macOS). Each result shows the page name and an excerpt with the matching words highlighted.
Reopening search brings back your last query and its results in that browser tab, selected so you can type over it.
Like the Page Picker, search has three segments:
The dropdown next to the input chooses how your query is matched:
All words and Exact phrase handle these languages fully:
bikes finds bike).مكتبة finds المكتبة), and both modes ignore how it is spelled: alef with or without hamza or madda, ة or ه, ى or ي, short vowel marks and tatweel.東京 finds 東京タワー.In languages written in the Latin alphabet, a word typed without its accents also finds it with them (zolw finds żółw, cafe finds café). Pages with the accents as you typed them come first.
In other languages, such as Polish, Russian or Turkish, search finds words only in the form you typed: żółwie doesn’t find żółw. In spaces with semantic search, Semantic finds them in any form.
Search covers the Markdown pages in your space, up to 1 MiB each. Built-in library pages bundled with SilverBullet and attachments are not searched.
The server keeps a full-text index of the space and updates it as pages change, usually within a second. After the server starts, it builds or catches up the index in the background. Until that finishes, results come with a “Search index building” notice and may be incomplete.
Where the index lives:
search/indexes/ in the server’s data folder, so restarts reuse it.silverbullet publish, in memory only, rebuilt each time the server starts. Such a site has no semantic search. See Publishing#Search.When enabled (default off), SilverBullet offers semantic search — search based on meaning. It runs an embedding model on the server itself (in Desktop, on your computer). Beyond downloading the model, no data leaves your server/desktop. On a server, an administrator turns it on in the dashboard. Desktop is covered under #SilverBullet Desktop.
After selecting a model, it is downloaded once. Choosing a model turns semantic search on for every space by default. The dashboard asks first, saying how many spaces will build a semantic index.
Once the semantic index is ready, a Related Pages panel becomes available (toggle it with Navigate: Related Pages). It lists up to 5 content pages, each with a formatted excerpt from its closest matching passage. From Reference/Space Lua, search.related() returns the same list (see API/search).
granite-embedding-97m-multilingual-r2 (default): 52 languages (and 200+ with less training), a 220 MB download and about 490 MB of memory.multilingual-e5-small: About 100 languages, a 488 MB download and about 590 MB of memory.Granite Embedding 97M Multilingual R2 (IBM, Apache-2.0) is the recommended model and comes first in the list. It finds paraphrases and pages in other languages better. Multilingual E5 Small (MIT) stays available, and a server, space or Desktop already set to it keeps using it. Switching to another model builds new semantic indexes.
Semantic search is good at paraphrases, questions and other languages. For names, codes and exact wording, All words and Exact phrase stay the better choice.
Once a model is configured, spaces with up to 5,000 Markdown pages are eligible for semantic search by default. Larger spaces require explicit opt-in. A space that grows past this threshold keeps its existing semantic search and shows a warning about relevance. 5,000 is a default policy threshold, not a hard page limit. Full-text search remains available.
Open Settings → Advanced and choose a model under Semantic search. The download starts right away. Semantic search is then on for every space. To keep a space on full-text search only, open its settings, go to Advanced and uncheck Semantic search. The same 5,000-page rule applies. Changing the model reloads the open spaces, after asking first. The model and the indexes live in the app’s data folder.
The same search is available from Reference/Space Lua as search.status() and search.find(). See API/search.
From the command line, run it with sb lua (see Integrations/CLI):
sb lua 'search.find("garden plan")'
sb lua 'search.find("flat tire", {mode = "phrase", limit = 10})'
The result is JSON with items (each with page, line, excerpt and more) and total.