Finding things
By the end of this you will know how to search a workspace, how to narrow a search to one part of it, and why the workspace's language decides what comes back.
Search a workspace
Every folder page carries a search box. From the workspace root it searches everything; from inside a folder it searches that folder and everything under it.
A query matches a document's title and its body. A match in the title ranks above a match only in the body, because a document named for the thing you asked about is usually the one you wanted. Results are ordered by rank, then by most recently updated, so the same query twice gives the same order rather than an arbitrary one.
Each result shows the document's full path and a short excerpt with the matched terms marked in context. The path matters once folders are in use: a name alone no longer identifies a document.
An empty query returns nothing. It is not an error, and it deliberately does not return everything.
Scope it to a folder
Searching from inside bewertung covers bewertung and everything beneath it,
and nothing else. This is the same idea as pointing a connected client at one
workspace, one level further down — useful when a knowledge base has grown
several distinct areas and a query for a common word would otherwise pull in all
of them.
The same option exists over MCP: search_documents takes an optional folder,
so a client can be aimed at one part of a knowledge base.
Language decides what you find
Search stems in the workspace's configured language, and the default is German.
This is not cosmetic. German is heavily inflected and compound-heavy, so a
search for Bewertung has to find Bewertungen, and one that does not is not a
search. The same configuration builds the index and parses the query — a
mismatch would return quietly wrong results rather than failing, which is why it
is a property of the workspace rather than of the query.
A workspace can be set to English or to simple instead. Changing it needs the
index rebuilt, which is a projection rebuild and therefore routine:
pnpm --filter @md-platform/core rebuild-projection <workspaceId>
What search will not do
- It will not cross a workspace. Search is scoped from your credentials, and there is no query, parameter, or tool argument that widens it. A term that appears only in another workspace's documents returns nothing at all — not a count, not a redacted row.
- It will not find an uploaded file's contents. Uploads are indexed by name and type, not by what is inside the PDF.
- It is not a semantic search. It matches words, with stemming. Asking it a question in a sentence works less well than searching for the terms you expect the document to contain — which is one of several reasons how you write changes what both readers can find.
What "done" looks like
A query from the workspace root finds a document by a word in its body; the same query from a sibling folder does not.
Next: connecting an AI client.