Pure, stateless scoring + ranking for the global settings search. More...
#include <PhosphorControl/SearchRanker.h>
Static Public Member Functions | |
| static int | score (const QString &query, const SearchEntry &entry) |
| Score entry against query (case-insensitive). | |
| static QVector< SearchEntry > | rank (const QString &query, const QVector< SearchEntry > &entries, int limit=-1) |
| Return entries with score > 0, sorted by score desc (stable: input order preserved on ties). | |
| static int | editDistance (const QString &a, const QString &b) |
| Classic Levenshtein edit distance over foldForSearch-normalised inputs, so it is case-, accent- and ß-insensitive. | |
| static QString | closestTitle (const QString &query, const QVector< SearchEntry > &entries, int maxDistance=-1) |
| Title of the entry whose title is the closest edit-distance match to query, for a "did you mean …" hint. | |
| static void | prefold (SearchEntry &entry) |
Fill entry's folded* fields from its display strings. | |
| static QString | foldForSearch (const QString &s) |
| Normalise a string for search comparison: decompose, drop combining marks, then simple case-fold, with ß and ẞ mapped to "ss" explicitly afterwards. | |
Pure, stateless scoring + ranking for the global settings search.
Typo-tolerant ranking WITHOUT leading with edit distance (which surfaces garbage on short queries). Per field, the best tier wins; fields are weighted title > keywords > subtitle:
exact (1000) > prefix (800) > word-start prefix (600) > substring (400) > subsequence (120–330, denser = higher)
A score of 0 means "no match" (filtered out). Edit distance is offered separately as closestTitle() for a "did you mean …" fallback that callers use ONLY when rank() returns nothing.
Stateless + Qt-free beyond QString — fully unit-testable headless.
|
static |
Title of the entry whose title is the closest edit-distance match to query, for a "did you mean …" hint.
Returns empty when nothing is within maxDistance (default scales with query length). Intended for the zero-results case only.
|
static |
Classic Levenshtein edit distance over foldForSearch-normalised inputs, so it is case-, accent- and ß-insensitive.
Retained as exported API; no in-tree production caller (closestTitle uses the pre-folded variant so it folds each side once).
|
static |
Normalise a string for search comparison: decompose, drop combining marks, then simple case-fold, with ß and ẞ mapped to "ss" explicitly afterwards.
So "Café" matches "cafe" and "Größe" matches "grosse".
The explicit mapping is NOT redundant: Qt's toCaseFolded() is a SIMPLE fold, so it maps ẞ (U+1E9E) down to ß and never to "ss". Deleting the special case would silently break exactly the strings it names.
LOSSY BY DESIGN where a combining mark carries meaning: Vietnamese tone marks collapse, so má / mà / mả all fold to "ma" and such a query matches more titles than it names. Both sides fold identically, so that case only ever WIDENS the result set rather than hiding a match.
It does NOT fold characters with no canonical decomposition: ø, ł, đ and ı survive, so "Størrelse" is not reachable by typing "storrelse". That direction NARROWS, unlike the combining-mark case above. ß is special-cased because it is the one such character common enough in settings text to matter.
Returns an EMPTY string for input that is entirely combining marks. Callers matching with contains() must guard that case: contains("") is true, which would match everything.
Exposed rather than kept file-local because the sidebar rail (SidebarRows) matches its own breadcrumbs while this class matches the global index. Both must fold identically or the same query returns different results in the two surfaces, which reads as a bug in whichever one the user tried second.
|
static |
Fill entry's folded* fields from its display strings.
Call once per entry when the index is built: rank() then matches against the folded copies instead of folding every field of every entry on every keystroke. Idempotent.
|
static |
Return entries with score > 0, sorted by score desc (stable: input order preserved on ties).
limit < 0 = no cap.
|
static |
Score entry against query (case-insensitive).
0 = no match.
Retained as exported API and as the tested reference implementation of the tiering; it has no in-tree production caller (rank() is what the controller uses). Folds query on EVERY call, so it is a single-entry probe, not a loop body. rank() is the loop-safe entry point: it folds the query once and matches against each entry's pre-folded fields. Adopting this in a loop would reintroduce the per-entry folding cost that rank() avoids.