Compare two documents by meaning
See which paragraphs say the same thing in other words, which changed meaning, and which exist in only one version. A language model runs on your device. No API key, and your documents stay in the browser.
The first comparison downloads a 90 MB language model once; your browser keeps it for next time.
Saved comparisons
Sign in to keep comparison reports on your SkillSafe account and reopen them on any device. Optional and free.
How it works
- Each document is split into paragraphs (or sentences). Long paragraphs are compared in parts of about 90 words, the most the model reads at once.
- all-MiniLM-L6-v2, a sentence-embedding model, turns every part into a vector on your device. Cosine similarity between vectors measures closeness in meaning, from 0 (unrelated) to 1 (same).
- Identical paragraphs pair first; the rest pair one to one from the most similar down, above the match floor. Pairs out of document order are marked as moved.
- Embeddings score "hired" and "fired" as close, so every pair is also checked for added or removed negation, changed numbers and dates, changed obligation words (must, should, may), changed scope words, and opposite words. Any hit marks the pair as changed meaning, with the reason.
The model is trained on English text. Other languages compare less reliably.
Questions
How is this different from a normal diff or Word's Compare?
A normal diff compares characters, so a reworded paragraph shows as a wall of red and green and a moved paragraph shows as deleted and re-added. This tool matches paragraphs by meaning first, so a reworded paragraph is marked as the same point in other words, a moved one is marked as moved, and only the paragraphs whose meaning changed, or that exist in one version only, stand out.
Can it tell when a small edit reverses the meaning?
Partly, and it says how. Sentence embeddings rate 'may sublet' and 'may not sublet' as very similar, so the tool also checks every matched pair for added or removed negation, changed numbers and dates, changed obligation words such as must, should and may, changed scope words, and opposite words such as hired and fired. A pair that trips any check is marked changed meaning with the reason shown. It cannot catch every reversal, so read the changed and reworded pairs that matter.
Do I need an OpenAI API key or an account?
No. The model runs in your browser with WebAssembly or WebGPU. There is no API key, no credit cost and no sign-in. The model is a one-time 90 MB download that your browser keeps. Signing in is optional and only saves comparison reports to your account.
Are my documents uploaded?
No. Files are read and compared inside the browser tab. Only the model weights are downloaded.
Which languages and file types work?
The model is trained on English, so similarity scores are reliable for English text. PDF files with a text layer, Word DOCX, OpenDocument, RTF, HTML, EPUB, Markdown and plain text can be opened; scanned PDFs without a text layer cannot.