Skip to content

suggestions.md

Latest commit

 

History

History
75 lines (57 loc) · 3.99 KB

suggestions.md

File metadata and controls

75 lines (57 loc) · 3.99 KB
Raw

Suggestions & dictionaries

The smartbar shows up to SuggestionCount candidates from the active language's ISuggestionProvider.

What the default provider does

Feature Details
Completion Most frequent dictionary words starting with the typed prefix (case and diacritic insensitive: cafe → café).
Correction Bounded Damerau-Levenshtein search in a trie. Substitutions of physically adjacent keys and transpositions are cheaper, so typos like teh or hwllo rank well.
Auto-correct When the typed word is unknown (or a rare spelling one edit away from a much more frequent word), the best correction is highlighted and applied when you type a space or punctuation. Backspace right after it restores what you typed (UndoAutoCorrectOnBackspace) and the word is not corrected again.
Typed word Correctly spelled words are shown first; unknown words are offered so they can be kept.
Learning Committed words and word pairs are learned (LearnWords, disabled in incognito mode, password and private fields). New words are suggested after being used twice (UserDictionary.LearnThreshold).
Next-word prediction After a word, the most frequent learned follow-up words are offered (NextWordPrediction).
Removing Long press a learned suggestion to remove and block it.
Emoji :pizza style shortcodes suggest emojis, and words matching an emoji name add one emoji candidate (EmojiSuggestions).
Clipboard Text copied within the last minute is offered once (ClipboardSuggestion).

Related typing features: auto-capitalization at sentence start (respecting the field's capitalization mode), double-space period, and smart punctuation spacing (word . → word. after a suggestion was committed).

Dictionaries

English (~50,000 words with frequencies, from FlorisBoard) is bundled. Register other languages with a WordDictionary — FlorisBoard JSON ({"word": frequency}), plain text (word [frequency] per line) or pairs:

var models = keyboard.Engine.LanguageModels;
models.Register("de", () => WordDictionary.FromText(File.ReadAllText("de_wordlist.txt"), "de"));
models.Register("fr", () => WordDictionary.FromJson(File.ReadAllText("fr.json"), "fr"));
models.Register("xx", WordDictionary.FromWords([("hello", 200), ("world", 180)], "xx"));

Frequencies are on a 0–255 scale (higher is more frequent). Dictionaries load lazily on first use. Glide typing uses the same dictionaries.

User dictionaries

var user = keyboard.Engine.LanguageModels.GetUserDictionary("en");
string json = user.Export();   // persist
user.Import(json);             // restore
user.Block("damn");            // never suggest
user.Clear();                  // forget learned words
keyboard.Engine.LanguageModels.UserDictionaryChanged += (_, language) => Save(language);

Custom providers

Plug in any engine (a server side language model, a domain vocabulary, ...):

public sealed class ProductNames : ISuggestionProvider
{
    public IReadOnlyList<Suggestion> Suggest(SuggestionRequest request) =>
        _catalog.Where(p => p.StartsWith(request.Word, StringComparison.OrdinalIgnoreCase))
                .Take(request.MaxCount)
                .Select(p => new Suggestion(p, SuggestionKind.Completion, 0.9))
                .ToList();

    public void Learn(string word, string? previousWord) { }
    public void Forget(string word) { }
}

keyboard.Engine.SuggestionProviderFactory = subtype => subtype.LanguageTag.StartsWith("en") ? new ProductNames() : null;

Return IsAutoCommit = true on a suggestion to have it applied on the next separator. SuggestionRequest.KeyProximity provides the key-distance cost function of the current layout.

Controlling suggestions per field

  • TextBox.IsTextPredictionEnabled="False" or sk:ScreenKeyboardInput.SuggestionsEnabled="False" disables them.
  • Password fields never show suggestions and never learn.
  • sk:ScreenKeyboardInput.IsPrivate="True" keeps suggestions but disables learning and history.