"""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)