Text Diff Checker

Why patches include context

A diff records an edit rather than reproducing two complete documents. GNU’s unified-format documentation describes the compact unified format: changed lines appear with nearby context instead of duplicating all surrounding text.

Context helps a reviewer locate the edit. The downloaded patch names original.txt and revised.txt and includes line-based changes even when the on-screen view compares words or characters.

A repeated line may have more than one plausible alignment. The engine finds an edit sequence; it does not understand author intent. Review surrounding text before treating a moved sentence as a deletion or an addition.

Visually equal is not always identical

An accented letter can be one code point or a base letter followed by a combining mark. The Unicode normalization specification defines canonical normalization. The optional NFC control compares canonically equivalent sequences without changing the editors.

CRLF and LF are distinct line-ending sequences. Exact comparison keeps that difference; Normalize line endings treats CRLF and lone CR as LF. Ignoring blank lines removes empty separators, including their boundary information.

Similarity is twice the unchanged token count divided by both token totals. Switching token types or ignore rules changes that percentage. It is not a plagiarism score, a semantic similarity model or evidence of correctness.

Choose the unit that answers your review question

Line mode treats a complete line, including its ending, as a token. It is good for lists, log excerpts, outlines, subtitles and configuration, but replacing one word can mark the whole line as changed. Word mode follows the comparison engine's tokenization of text, punctuation and significant spacing. Depending on the script and punctuation, a natural word may be split into fragments; the totals count comparison tokens, not linguistic words. Character mode helps inspect a decimal point, URL, identifier or short spelling change.

Start with structure, then narrow the unit for the uncertain edit. A paragraph moved elsewhere can appear as a removal and an addition, not a recognized move. Repeated identical lines allow more than one reasonable alignment, so follow surrounding context. This tool compares two pasted strings; it has no repository history, three-way merge, tracked comments or understanding of why the author revised them.

Ignore rules are decisions about meaning

All ignore and normalization rules start off. Turn one on only after seeing what the exact comparison reports. Ignore case uses lowercase comparison, not a language-aware collation or a rule for equivalent identifiers. Ignore blank lines removes whitespace-only lines and rejoins the remainder with LF; it can therefore also hide line-ending and final-boundary differences. It is not merely a visual collapse of empty rows.

Whitespace behaves differently by unit. In Lines, trimming and collapsing spaces or tabs can hide indentation and even an ending-only change when that rule is enabled. In Words, whitespace runs become single spaces before comparison. In Characters, whitespace is removed altogether: a b and ab can then match. Never enable this rule just to improve the percentage; decide whether the omitted distinctions can affect your task.

Read the result without overreading the score

The original is on the left and the revision on the right. Minus/red denotes removal; plus/green denotes addition. Neutral content matched under the current rules, not necessarily byte-for-byte. Split view keeps the two sides separate; unified view interleaves the sequence and is convenient for a compact review. Swap reverses the direction, so previous additions become removals. It does not decide which version is better.

For a reproducible example, use Lines with no ignore rules: original limit = 8, revision limit = 12, each followed by a newline and the identical line retry = true with its newline. There is one removed line, one added line and one unchanged line. Similarity is 50%: 2 × 1 / (2 + 2). Word or character mode counts different tokens and yields a different score for the same edit.

The exported patch and the screen have different jobs

Copy Diff and Download Diff both use a unified, line-based patch of the raw inputs. The files are labelled original.txt and revised.txt. Ignore rules, NFC and newline normalization do not rewrite that patch. Thus a normalized screen can show no differences while the export still records a real representation change. Keep the settings alongside the patch if another reviewer needs to understand that distinction.

The header identifies before and after files; a hunk header identifies the affected line ranges. Within a hunk, a leading space is context, a minus is an old line and a plus is a new line. Missing final newlines can be recorded too. A patch is a recipe, not a saved repository commit. Before applying one elsewhere, confirm the destination and baseline in the target tool; this page does not apply patches or resolve conflicts.

Diagnose surprising results in small steps

A comparison error means the computation did not complete, not that the texts are identical. Each input is limited to 50,000 UTF-16 string units, and difficult edit sequences can exceed the work budget even below that size. Patch generation has its own bounded calculation and can also be unavailable. Split large texts at meaningful section boundaries while retaining enough context to locate the edits; do not judge an error-state percentage as a result.

For invisible differences, verify that both pastes are complete, disable all rules, then compare a short segment in Characters. NFC does not equate straight and curly quotes, ordinary and nonbreaking spaces, or visually similar Latin and Cyrillic letters. Character tokens are code points, not entire emoji or grapheme clusters. For formatting, comments, tables or binary encodings, return to the original document or a file-aware comparison tool rather than expecting pasted text to preserve them.