Document Converter

Three engines, chosen by what the document is

Document conversion is not one process. Which engine runs depends on what you are converting and what you want out, and the differences are worth knowing because they explain the results you get.

LibreOffice handles office documents — Word, Excel, PowerPoint, OpenDocument — by opening them headlessly and exporting through a filter. It has three separate filters, Writer, Calc, and Impress, and each one has its own list of permitted outputs. That is why a spreadsheet cannot be exported directly to a presentation format: LibreOffice will not convert across those categories, so conversions that need to cross them route through an intermediate step instead.

pandoc handles structure-first conversions: Markdown, HTML, EPUB, and plain text, in any combination. It parses the source into a document model — headings, lists, emphasis, links — and writes the target from that model. Because it works on structure rather than appearance, the output is a real document rather than a rendering of one.

pdf2docx handles the hard direction: turning a PDF back into an editable Word file, which is an inference problem rather than a translation.

What the conversion will and will not preserve

Text, headings, lists, tables, and basic styling carry across reliably. The things that go wrong are almost always about appearance rather than content.

• Fonts. The server does not have the fonts installed on your machine. Missing fonts are detected and substituted before rendering, but a substitute has different metrics, which moves line breaks and can change your page count.
• Right-to-left layout. Arabic documents get an extra pass: justification is normalized before rendering so paragraphs do not come out stretched, and a font known to render badly headlessly is swapped for one that does not.
• Complex layout. Multi-column pages, floating figures, and text boxes are the cases where a converted document needs cleanup.
• Formulas in spreadsheets. They export as their last calculated value, not as live formulas.
• Animations and transitions in presentations. These have no equivalent outside a live deck.

Structured data counts as a document here

JSON, XML, YAML, and TOML convert into readable documents and spreadsheets: nested keys and arrays are flattened into a shape a person can read, and written as PDF, DOCX, CSV, or plain text. An array of flat records converts to a clean table; a deeply nested configuration tree converts to a readable rendering of the tree, because there is no meaningful table to make from it.

Malformed input stops the job with a parse error rather than producing a half-converted document — a converter that silently skipped what it could not read would hand you something that looks complete and is missing records.