CodeKeyboard Part 6: Fuzzy Correction, Emoji, Snippets, Trie Decay

Spell correction / fuzzy fill

ADR vs code

docs/architecture/decisions/ADR-005-spell-correction.md evaluates:

• Schulz & Mihov Levenshtein automata
• Hanov 2011 DP-row DFS on a trie
• BEVA (Zhou et al. 2016)
• SymSpell (rejected: memory)
• BK-trees (rejected)

ADR accepts Hanov as v1. Production path uses BEVA.

MergedSuggestionStrategy.fuzzyFill calls BevaTrieSearch.search on UserTrieAdapter and BaseTrieAdapter.
FuzzyTrieSearch.kt still implements Hanov DFS and is used in tests/benchmarks.

When fuzzy runs

val exact = exactSuggest(prefix, k)
if (exact.size >= k) return exact
val threshold = FuzzyThreshold.forLength(prefix.length)
if (threshold == 0) return exact
// fill k - exact.size

// FuzzyThreshold
len <= 3 -> 0
len == 4 -> 1
else     -> 2

So fuzzy is not “only when exact is empty”; it fills remaining slots. Short prefixes skip fuzzy entirely.

Ranking rule

Comment in Merged: collect all within threshold (Int.MAX_VALUE cap on search), then sort by edit distance, then longer common prefix with query, then frequency. Do not stop DFS early at maxResults if that yields alphabetical junk (historical bug).

BEVA sketch (from file header): per-node edit-vector bitmasks evBits[e]; transitions for match/sub/del/ins; prune when no reachable bits; terminal edit distance = min e with bit n set. Words up to length 30 fit in 32-bit masks.

Emoji panel

Data pipeline

scripts/gen_emoji.py:

• Source URL Unicode 15.0 emoji-test.txt (cached under scripts/)
• fully-qualified only
• skin tones stripped into variants under base
• writes android/app/src/main/assets/emoji.json

Measured size: 171457 bytes.

UI

EmojiPanelView.kt, not SuggestionBarView.

IME:

// showEmojiPanel: build with kbHeight + nav padding, setInputView(emojiPanel)
// hideEmojiPanel: emojiPanel = null; setInputView(onCreateInputView())

Tap emoji: currentInputConnection?.commitText(emoji, 1).

Focus (prevents IME dismiss):

descendantFocusability = FOCUS_BLOCK_DESCENDANTS
isFocusable = false

Category tab icons in companion CATEGORY_ICONS (Smileys, People, Animals, Food, Travel, Activities, Objects, Symbols, Flags, Component).

Height: onMeasure can lock to fixedHeightPx so the panel matches the keyboard slot; nav padding keeps content above system chrome. Several git commits were about panel height and tab layout.

Snippets

SnippetStore object, SharedPreferences via KeyboardSettings, key prefix snippet_.

First run seeds empty: em, ph, addr, me, gh, li when snippets_seeded is false.

fun matching(prefix: String): List<String>  // expansions, max 3 non-empty
fun add(shortcode, expansion): Boolean     // validates ^[a-z0-9_]+$, no collide
fun update / delete / get / exists / all

IME: typing ; starts composing with semicolon; suggestions from SnippetStore.matching instead of tries. Settings UI in RN writes the same prefs through the bridge.

User trie overnight decay

Problem: unbounded user.trie after months of inserts.

MainApplication.scheduleTrieDecay:

Constraints.Builder()
    .setRequiresCharging(true)
    .setRequiresDeviceIdle(true)
    .build()
PeriodicWorkRequestBuilder<TrieDecayWorker>(1, TimeUnit.DAYS)
WorkManager.enqueueUniquePeriodicWork("trie_decay", KEEP, request)

TrieDecayWorker:

val trie = UserTrie.load(file)
trie.applyDecay(factor = 0.9, newEpoch = trie.decayEpoch + 1)
trie.save(file)

UserTrie.applyDecay: multiply terminal freqs by 0.9^deltaEpoch, compact dead subtrees, recompute max freq, if totalNodes > 50_000 rebuild keeping top 5000 words by frequency.

Typing path does not call decay. Session end only saves the trie as-is.

ADR-004 describes the same constraints and hard cap idea.

Verification (part 6)