feat(desktop): built-in Han/Eng layout auto-switching, terminal/Vim support, and Unicode input

This commit is contained in:
Yun Chan 2026-10-27 10:48:07 +09:00
parent 1da04b9e46
commit a3ac44da2b
109 changed files with 8234 additions and 242 deletions

View file

@ -0,0 +1,78 @@
"""Hybrid: LLM first word vs personal n-gram guess on identical cases.
Reports exact first-word hits and 'useful' hits (guess is a prefix of the true first word: Korean stem + ending).
"""
import json
import re
import sqlite3
import sys
from collections import Counter, defaultdict
db_path, cases_path, out_path = sys.argv[1], sys.argv[2], sys.argv[3]
db = sqlite3.connect('file:' + db_path + '?mode=ro', uri=True)
texts = [r[0] for r in db.execute("select text from typing_samples where source in ('typed','voice')")]
texts += [r[0] for r in db.execute(
"select coalesce(polished_text, original_text) from history where coalesce(polished_text, original_text) is not null")]
cases = json.load(open(cases_path, encoding='utf-8'))
out = open(out_path, 'w', encoding='utf-8')
TRAIL = re.compile(r'[\s.!?。!?…~,"\')]+$')
def words(text):
return [TRAIL.sub('', w) for w in re.sub(r'\s+', ' ', text or '').strip().split(' ') if w]
def ngram_guess(prefix_words, exclude):
bi, tri = defaultdict(Counter), defaultdict(Counter)
for text in texts:
if exclude in re.sub(r'\s+', ' ', text):
continue
ws = words(text)
for i in range(1, len(ws)):
bi[ws[i - 1]][ws[i]] += 1
if i >= 2:
tri[(ws[i - 2], ws[i - 1])][ws[i]] += 1
if len(prefix_words) >= 2 and tri.get((prefix_words[-2], prefix_words[-1])):
word, n = tri[(prefix_words[-2], prefix_words[-1])].most_common(1)[0]
return word, n, 'tri'
if prefix_words and bi.get(prefix_words[-1]):
word, n = bi[prefix_words[-1]].most_common(1)[0]
return word, n, 'bi'
return None, 0, None
stats = Counter()
for c in cases:
truth = words(c['truth'])
if not truth:
continue
t0 = truth[0]
full = re.sub(r'\s+', ' ', c['prefix'] + c['truth']).strip()
g, n, kind = ngram_guess(words(c['prefix']), full)
llm = words(c['suggestion'])
l0 = llm[0] if llm else None
stats['cases'] += 1
for name, guess in (('llm', l0), ('ngram', g)):
if guess:
stats[f'{name}_shown'] += 1
stats[f'{name}_exact'] += guess == t0
stats[f'{name}_useful'] += len(guess) >= 1 and t0.startswith(guess)
if g and l0:
stats['both'] += 1
if g == l0 or l0.startswith(g):
stats['agree'] += 1
stats['agree_exact'] += g == t0
stats['agree_useful'] += t0.startswith(g)
# hybrid policy: ngram when it has a guess, otherwise LLM
h = g or l0
if h:
stats['hyb_shown'] += 1
stats['hyb_useful'] += t0.startswith(h)
stats['hyb_exact'] += h == t0
N = stats['cases']
p = lambda k, d: f"{stats[k]}/{stats[d]} = {stats[k]/stats[d]:.3f}" if stats[d] else '-'
print(f'cases {N}', file=out)
for name in ('llm', 'ngram', 'hyb'):
print(f"{name:6} shown {stats[name + '_shown']/N:.3f} exact|shown {p(name + '_exact', name + '_shown')} useful|shown {p(name + '_useful', name + '_shown')}", file=out)
print(f"agree (ngram==LLM first word) {p('agree', 'both')} exact {p('agree_exact', 'agree')} useful {p('agree_useful', 'agree')}", file=out)

View file

@ -0,0 +1,63 @@
"""Personal n-gram next-word predictor on the same replay cases (leave-one-out).
Compares with the LLM on identical cases: does the user's own history predict the next word better?
"""
import json
import re
import sqlite3
import sys
from collections import Counter, defaultdict
db_path, cases_path, out_path = sys.argv[1], sys.argv[2], sys.argv[3]
db = sqlite3.connect('file:' + db_path + '?mode=ro', uri=True)
texts = [r[0] for r in db.execute("select text from typing_samples where source in ('typed','voice')")]
texts += [r[0] for r in db.execute("select coalesce(polished_text, original_text) from history where coalesce(polished_text, original_text) is not null")] if db.execute("select count(*) from sqlite_master where name='history'").fetchone()[0] else []
cases = json.load(open(cases_path, encoding='utf-8'))
out = open(out_path, 'w', encoding='utf-8')
TRAIL = re.compile(r'[\s.!?。!?…~,]+$')
def words(text):
return [w for w in re.sub(r'\s+', ' ', text).strip().split(' ') if w]
def build(exclude):
bi, tri = defaultdict(Counter), defaultdict(Counter)
for text in texts:
if exclude and exclude in re.sub(r'\s+', ' ', text):
continue
ws = [TRAIL.sub('', w) for w in words(text)]
for i in range(1, len(ws)):
bi[ws[i - 1]][ws[i]] += 1
if i >= 2:
tri[(ws[i - 2], ws[i - 1])][ws[i]] += 1
return bi, tri
shown = hits = 0
min_count = int(sys.argv[4]) if len(sys.argv) > 4 else 2
examples = []
for c in cases:
full = re.sub(r'\s+', ' ', (c['prefix'] + c['truth'])).strip()
bi, tri = build(full)
pw = [TRAIL.sub('', w) for w in words(c['prefix'])]
truth_first = TRAIL.sub('', words(c['truth'])[0]) if words(c['truth']) else ''
guess = None
if len(pw) >= 2 and tri.get((pw[-2], pw[-1])):
word, n = tri[(pw[-2], pw[-1])].most_common(1)[0]
if n >= min_count:
guess = word
if guess is None and pw and bi.get(pw[-1]):
word, n = bi[pw[-1]].most_common(1)[0]
if n >= min_count:
guess = word
if guess:
shown += 1
hit = guess == truth_first
hits += hit
if len(examples) < 20:
examples.append((c['prefix'][-25:], guess, truth_first, hit))
print(f'corpus texts {len(texts)}; cases {len(cases)}; min_count {min_count}', file=out)
print(f'ngram coverage {shown/len(cases):.3f} hit|shown {hits/shown if shown else 0:.3f} hits {hits}', file=out)
for e in examples:
print(' ', e, file=out)

View file

@ -0,0 +1,74 @@
import json
import math
import re
import sys
cases = json.load(open(sys.argv[1], encoding='utf-8'))
out = open(sys.argv[2], 'w', encoding='utf-8')
END = re.compile(r'[.!?。!?…\n]')
TRAIL = re.compile(r'[\s.!?。!?…~,]+$')
def norm(text):
return re.sub(r'\s+', ' ', text).strip()
def truncate(tokens, tau):
"""Keep tokens while each logprob >= tau; stop at sentence end. Drop a trailing partial word."""
text = ''
first_word_lp = 0.0
first_word_done = False
started = False
for tok in tokens:
t, lp = tok['token'], tok['logprob']
if '<' in t:
break
if lp < tau:
# partial word cut: if this token continues a word (no leading space), drop the tail of the last word
if text and not t.startswith(' ') and not text.endswith(' '):
text = text[: text.rfind(' ') + 1] if ' ' in text.strip() else ''
break
text += t
if text.strip():
started = True
if started and not first_word_done:
first_word_lp += lp
if ' ' in text.strip():
first_word_done = True
if END.search(t):
break
return norm(text), first_word_lp
def score(sugg, truth):
truth = norm(truth)
sw, tw = sugg.split(' ') if sugg else [], truth.split(' ')
m = 0
while m < min(len(sw), len(tw)) and TRAIL.sub('', sw[m]) == TRAIL.sub('', tw[m]):
m += 1
bare = TRAIL.sub('', sugg)
full = bool(bare) and truth.startswith(bare)
acc = bare if full else ' '.join(sw[:m])
return m, full, max(0, len(acc) - 1) if acc else 0
total_truth = sum(len(norm(c['truth'])) for c in cases)
print('cases', len(cases), file=out)
print('tau gate coverage hit|shown full|shown savedRate meanLen', file=out)
for tau in (-0.5, -1.0, -1.5, -2.0, -3.0, -99):
for gate in (-0.1, -0.3, -0.5, -1.0, -99):
shown = hits = full_n = saved = length = 0
for c in cases:
if not c['tokens']:
continue
sugg, fw = truncate(c['tokens'], tau)
if not sugg or fw < gate:
continue
shown += 1
m, full, s = score(sugg, c['truth'])
hits += m >= 1
full_n += full
saved += s
length += len(sugg)
if shown:
print(f'{tau:5} {gate:5} {shown/len(cases):.3f} {hits/shown:.3f} {full_n/shown:.3f} {saved/total_truth:.4f} {length/shown:.1f}', file=out)