Convert a Word Document to Text for an LLM

A .docx is a zip file full of XML. Every word you typed is in there, and so is a great deal that is not words: the table you built, the order you put things in, the heading that told a reader where a section began. Getting a Word document into a prompt means getting the words out of that container, and the interesting question is what comes with them.

This guide covers converting a Word document to text for an LLM prompt: what survives the trip, what flattens, and how to send a folder of them at once. There is a spot to try it on your own files partway down. Nothing is uploaded to us; the extraction runs in your browser.

The table is the part worth protecting

Most of a Word document is prose, and prose survives anything. Tables do not. A three-column table is only readable while its cells stay apart and its rows stay whole, and that is exactly what a careless read gives up first.

report.docx
Quarterly Report
Revenue by region
[ Region | Q1   | Q2   ]
[ EMEA   | 1200 | 1350 ]

Quarterly Report
Revenue by region, in thousands.

| Region | Q1   | Q2   |
| EMEA   | 1200 | 1350 |
| APAC   | 980  | 1105 |

Each row stays on its own line with its cells in column order, so the model can still tell that 1350 is EMEA's second quarter and not APAC's first. Paragraphs keep the order you wrote them in, which matters more than it sounds: a document read out of order reads as a different argument.

Send the whole folder, not one file at a time

A Word document rarely travels alone. The draft has an appendix in a spreadsheet, a reference in a PDF, and a page of notes in markdown, and the question you want to ask is about all of them together.

1. 1

   Drop your documents

   A folder or a loose handful. Word documents, PDFs, spreadsheets, slide decks, and plain text or markdown all go in together.

2. 2

   Each one is opened in your browser

   The text is pulled out on your machine. If the default filter recognises the usual junk in the folder, it skips it before it reaches the bundle.

3. 3

   Copy one bundle

   A single document with each file labeled by name, plus a live token count so you know it fits before you paste it.

Try it on your own documents

Drop a Word file below, or a folder with a few mixed in. It runs in your browser, there is no account, and nothing is uploaded to us.

What flattens

Headings arrive as ordinary lines of text. The word Notes on its own line was a section heading in Word and reads as a stray line of prose once extracted, because the level that made it a heading is styling, and styling is not text.

In practice this costs less than it sounds, since a model infers structure from a short line followed by a paragraph about as well as you do. It costs something when the hierarchy is the question: if you need the model to know that four subsections sat under one heading and not another, say so in the prompt, or keep a markdown copy of the outline and drop that in too.

Images, charts, and text boxes are not prose either. A diagram that carries the point of a page will not carry it into the bundle, so describe it in the question instead.

Mind the token and cost budget

Documents add up faster than they look, and a folder of drafts and appendices can run past a model's context window without any single file looking large. FileConcat counts tokens as you add and remove files and can show what a bundle costs for a given model, so you can trim to the documents that carry the answer first.