Near-Duplicate Row Detector: Find Eval Rows That Differ Only in Formatting

Lowercasing is a display operation and it is not symmetric. The Greek capital sigma lowercases to a final sigma at the end of a word and to a medial sigma elsewhere, so a capitalized Greek word and the same word already in lowercase come out different and stay in two clusters. Case folding is defined for comparison: it sends both sigmas to one letter, it turns the German sharp s into a double s, and it is language-neutral, so the answer does not change with the language settings on your machine.

This page applies the published folding table one code point at a time and never calls a lowercase function, so no lowercase special case can interact with it. That is what keeps a Turkish dotted capital I from producing one answer on your machine and another on a colleague's.

The W3C character model says compatibility normalization should not be applied unless it was explicitly asked for, and that is sound advice for matching text on the open web, where a false match is silent and permanent. The trade here runs the other way. This page never edits your file. It prints the transform it used, shows every raw variant beside what it normalized to, and turns the step off with one click. A false merge costs you a glance at two rows; a missed merge is the thing you came here to find.

The step still carries its own warning, above, naming the cases it merges. If your set is full of chemical formulas or exponents, turn it off first.

Anything that needs to read meaning. Two rows that ask the same question in different words keep different keys here no matter which switches you set, because every step is a character transform. Paraphrase detection needs an embedding model, and this page loads none.

It also does not catch partial overlap, where one row quotes half of another. That is graded similarity, measured with n-gram overlap rather than exact key equality, and it belongs to a different rung.

No. The file is read, parsed, normalized, and grouped in this tab. There is no upload, no server call, and no analytics payload that carries a file name, a size, a row count, a column name, a key, or a cell value. The export is built in the browser from what is already on this page.

Because a threshold you can drag is a threshold nobody can reproduce. The measure here is binary by construction: two rows share a key or they do not, and the whole transform that produced the key is printed above the table and carried in the export. Anyone can redo it by hand and get the same clusters. A slider would replace that with a number whose meaning depends on which library computed it.