MeaningDiff — semantic diff for human text (private, offline)

identical ≥0.88 reworded 0.68–0.88 weak 0.42–0.68 moved added removed

Privacy: all processing happens locally in your browser (Transformers.js). No server.

Use semantic embeddings (recommended) Model: not loaded First run downloads ~45MB (multilingual-MiniLM) and is cached via browser cache/IndexedDB.

—

How it works

Sentences are split (handles Arabic ؟ ، ؛).
Each sentence is embedded via Xenova/multilingual-MiniLM-L12-v2 in-browser (Transformers.js, WebAssembly/WebGPU). Normalized embeddings → cosine similarity matrix.
Optimal sentence alignment via dynamic programming (Needleman-Wunsch, gap penalty) + moved-sentence detection (far-match ≥0.72).
If model unavailable, falls back to lexical Jaccard + 3-gram similarity — still useful offline.
Everything runs locally. Installable PWA, works offline after model is cached.

Tip: use for reviewing article revisions, legal/policy redlines, prompt version diffs, translation QA. Not a plagiarism detector — shows meaning change.