Converters

The converter pipeline lets Diffoniq compare file formats it cannot read itself — PDF documents, SQLite databases, image metadata — by running an external program that turns the file into plain text. The text output of both sides is then compared with the normal diff engine.

Some formats Diffoniq converts itself, without any program of yours: see Built-in converters below. Your own rules come first; the built-in ones apply where none of yours matches.

Rules are configured under Options › Converters… Each rule consists of:

Name Optional display name; shown in the converted: badge of the header. Without a name the file masks stand in.
File masks Semicolon-separated wildcard list, e.g. *.pdf;*.docx. Matching is case-insensitive and applies to the file name (also inside archives).
Command The external command line. %src% is replaced with the source file, %out% with the output text file the converter should write. A command without %out% has its standard output captured instead. Both paths are inserted already quoted — do not add your own quotation marks around the placeholders.
Timeout Maximum runtime in seconds (default 30). A converter that exceeds it is terminated, including any processes it started.
Active Deactivated rules stay in the list but are skipped.

The first active rule whose mask matches wins — use Move Up/Move Down to order overlapping rules. When no rule matches, the file is compared normally.

The fields below the list are moved into it with Add as New Rule or Apply to Selected. OK keeps what you typed as well: an entry that was never added becomes a new rule, and if a selected rule was changed in the fields, Diffoniq asks whether to apply the change to it or add it as a new rule. A rule the list already holds with the same fields is not added a second time. Cancel discards every change made in the dialog. The Help button (or F1) opens this page.

Built-in converters

Behind your own rules Diffoniq keeps converters of its own. They do not start another program: Diffoniq runs a second copy of itself, without a window, to read the file. A damaged file can therefore only end that helper, never the comparison window, and the timeout, the cache and the error hint work exactly as for your own rules. The helper may use at most 1 GB of memory; a file that would need more is compared unconverted, with a hint. A part that the file refers to but does not contain — a chapter, a slide, a header — appears as a line such as --- Slide 2 (missing) ---, so a comparison never reads as equal for something that was not there to read. The header badge reads, for example, converted: JSON (built-in).

An old Word file (.doc) against a new one (.docx), both converted by the built-in converters; three changes in the body text and one in the footnote are marked.
An old Word file against a new one: both read by the built-in converters, the footnote under its own heading.

Use the built-in converters (after your own rules) at the bottom of the dialog switches all of them on or off; they are on by default. To switch off a single one, list its name in diffoniq.ini, for example BuiltinConvertersOff=json (several separated by semicolons). A rule of your own for the same masks also takes precedence.

When a file one of them reads is compared unconverted because the switch is off, the header line says so instead of converted: not converted: built-in converters are off, or for a single one, for example, not converted: the built-in Word converter is off. The command line and the MCP server do not use this switch; their converters are always on.

ConverterMasksWhat the text shows
Word *.docx;*.docm;*.dotx;*.dotm One line per paragraph; a table row is one line with the cells separated by |. Tracked changes show the final state (insertions in, deletions out). Text boxes are included, field codes are not (their results are). Footnote and endnote references appear as [1]; headers, footers, footnotes and endnotes follow the body, each under its own --- heading. Comments, formatting, list numbers and pictures are not part of the text. A password-protected document cannot be read.
Word 97-2003 *.doc;*.dot The rules of Word for the old binary format: one line per paragraph, table rows with |, field results, footnote references as [1]; footnotes, endnotes, headers and footers and text boxes each under its own --- heading. Tracked changes are not resolved: text deleted with revision marks still shows. A password-protected document cannot be read; Word 6 and Word 95 files are not converted.
Excel 97-2003 *.xls;*.xlt The rules of Excel for the old binary format: every worksheet under --- Sheet: name ---, one line per row, empty rows in their place, formulas beside their values as in 4370 (=B4*C4), dates as dates. A function the converter does not know by name appears as FUNC123(...); a formula it cannot read is left out and the cell shows its value. A password-protected workbook cannot be read; Excel 5 and Excel 95 workbooks are not converted.
PowerPoint 97-2003 *.ppt;*.pot;*.pps The slides in the order of the presentation, each under --- Slide n ---: the title and body text first, then the text of the other shapes, one line per paragraph; the speaker notes follow under Notes:. Hidden slides are not marked. A password-protected presentation cannot be read.
Excel *.xlsx;*.xlsm;*.xltx;*.xltm Every worksheet in workbook order under --- Sheet: name --- (hidden sheets marked). One line per row with a value, cells separated by |; empty cells stay as gaps so the columns line up, and a value far to the right is written as [XFD] value. Empty rows stay as empty lines, so the rows keep their place under the sheet heading; after more than ten empty rows the next row starts with [row 500]. A formula stands beside its last calculated value, as in 4370 (=B4*C4): a changed formula is a difference even when its result stays the same. A file saved without calculated values shows the formula alone. Dates and times are written as 2026-11-09, 13:45:00 or both, whatever their cell format shows; other numbers exactly as Excel stored them. Charts, comments, formatting and chart sheets are not part of the text; binary .xlsb and old .xls files are not converted.
PowerPoint *.pptx;*.pptm;*.potx;*.potm;*.ppsx;*.ppsm The slides in the order of the presentation, each under --- Slide n --- (hidden slides marked), with the text of every shape, one line per paragraph, and table rows as one line with | between the cells. The speaker notes of a slide follow under Notes:. Layouts, masters, pictures, charts and formatting are not part of the text.
OpenDocument *.odt;*.ott;*.ods;*.ots;*.odp;*.otp The rules of Word, Excel and PowerPoint for text documents, spreadsheets and presentations of LibreOffice and other OpenDocument programs: tracked changes in their final state, footnotes under --- Footnotes ---, headers and footers under --- Header (page style) ---; sheets row by row with the stored values, formulas beside them and dates as dates (rows a sheet repeats with content are written at most 1000 times, followed by a line with the full count, empty ones not at all); slides with their speaker notes. Comments are not part of the text.
EPUB *.epub The chapters in reading order, each under --- file ---, one line per paragraph, heading or list item, table rows with | between the cells; white space as a browser shows it, preformatted text as it is. Pictures, styles and scripts are not part of the text. A copy-protected (DRM) book cannot be read.
JSON *.json;*.jsonc;*.geojson The data, not the layout: object keys sorted, two spaces indent, so key order, white space, comments (//, /* */) and a comma after the last element (as in settings.json of VS Code) are not compared. Numbers stay exactly as written — 1e5 and 100000 count as different. Strings are shown in one form, so "\u00e9" and "é" are equal. A key that occurs twice in an object is never dropped: every occurrence is listed with the note // duplicate key "a" (1 of 2). A file that is not valid JSON, or one larger than 32 MB, is compared as text, with the reason in the hint.

Your own rules act in the comparison window only. The built-in converters also serve a file comparison of the MCP server (compare_file) and of /report --convert; there they run with their factory settings, whatever this dialog says. Folder comparisons, /sync and /report without --convert compare the original bytes.

How converted comparisons behave

Caching

Conversion results are cached in the convcache folder next to diffoniq.ini (a portable installation keeps it portable). The cache key is the source file’s content plus the rule’s command line: an unchanged file never runs the converter twice, an edited rule never serves stale output. For a built-in converter the key holds its version instead of the command line, so a new Diffoniq version with an improved converter reads the file again. The entry ConvCacheMax=200 in diffoniq.ini caps the cache size in MB (least recently used entries are removed first; 0 disables the cap).

Security note

Diffoniq only ever runs commands you configured yourself — the built-in converters start nothing but Diffoniq itself, and nothing starts automatically for unmatched files. The command runs directly (no shell); shell syntax such as pipes or redirection only works when you explicitly configure cmd.exe /c as the command.

Recipes

Free, widely-used tools that work well; all examples write UTF-8, which Diffoniq detects automatically (UTF-16 and ANSI output are handled too). Word, Excel, PowerPoint, OpenDocument, EPUB and JSON need no recipe: the built-in converters read them, and a rule of your own for their masks would replace them.

PDF — pdftotext (Xpdf/Poppler) pdftotext -layout -enc UTF-8 %src% %out%
Masks: *.pdf
Rich Text — pandoc pandoc -t plain -o %out% %src%
Masks: *.rtf
SQLite databases — sqlite3 sqlite3 %src% .dump
Masks: *.db;*.sqlite;*.sqlite3
Image / photo metadata — ExifTool exiftool -charset UTF8 %src%
Masks: *.jpg;*.jpeg;*.tif;*.nef;*.cr2 — compares the metadata, not the pixels.

Tools that are not on the PATH need their full path, e.g. "C:\Program Files\pandoc\pandoc.exe" -t plain -o %out% %src%.

Trying it out without installing anything

Windows ships a tool that turns any file into text: cmd.exe /c certutil -encode -f %src% %out% writes the file as Base64. With the masks set to *.pdf, a PDF pair then opens in the converted, read-only text view with the converted: badge, and a different pair shows a difference. This is a way to see the pipeline work before installing a real converter — it is not a way to read the document: Base64 lines carry no words, and two PDFs that differ only in metadata differ in the text just as much as two with different content. For the document itself, use pdftotext from the table above.