Calculate line, word, or character similarity locally from equal versus inserted/deleted diff tokens.
Comparison runs locally in your browser. Interactive diff uses a bounded token count to avoid locking the page on extremely large inputs.
Compare texts with token counts, Jaccard overlap, length delta and a configurable similarity threshold alongside the primary diff score.
Use the same two primary text inputs to compare edit-based output with unique-token Jaccard and term-frequency cosine evidence. This remains lexical and does not claim semantic understanding.
The percentage is derived from the edit alignment at the selected line, word or grapheme granularity. Equal, inserted and deleted token counts remain visible so the score is inspectable rather than opaque.
Use the tool first, then use these checks to interpret, verify and hand off the result without displacing the primary workflow.
The primary diff score, Jaccard overlap and cosine term-frequency score answer different questions. Diff emphasizes edit operations, Jaccard compares unique token sets, and cosine uses token frequency. A high value under one model does not prove semantic equivalence, authorship, plagiarism, or factual agreement.
A score is easier to trust when you can see the tokens contributing to it. The supplemental panel lists frequently shared tokens so boilerplate, repeated names, or common vocabulary do not hide behind one percentage. For decisions about meaningful phrase reuse, inspect the actual passages rather than relying only on token overlap.
The supplemental word model lowercases Unicode letters and numbers. It does not remove stopwords, stem words, translate languages, or use embeddings. Those choices deliberately keep the evidence reproducible, but they also mean inflection, synonyms, punctuation structure and word order may be treated differently than a semantic model would.
A fixed cutoff such as 70 or 80 percent can be useful for triage, but the right threshold depends on text length, domain and consequences. Test known similar and known dissimilar examples from the real workflow, then document the metric and threshold together so another reviewer can reproduce the decision.
WebToolArc