abacusai/abacusai-bot/docx

Working with Word documents

Write a CommonJS script (make-doc.cjs) and run it with the app's own Node:

"$ABACUSAI_BOT_NODE" make-doc.cjs

require() finds these with nothing to install:

Reading

const mammoth = require("mammoth");
const { value } = await mammoth.extractRawText({ path: "in.docx" });

For headings, lists and tables, convert to HTML instead. The tags keep the structure, and paragraph styles show up as h1, h2 and so on:

const { value: html, messages } = await mammoth.convertToHtml({ path: "in.docx" });

messages lists what mammoth could not map. Read it before you trust the output. Wrap top-level await in (async () => { ... })(), since a .cjs script has none.

Writing

Use the built-in heading styles and real lists rather than manual formatting. They are what make a document look native, and what its table of contents and navigation pane read:

const fs = require("fs");
const { Document, Packer, Paragraph, HeadingLevel, Table, TableRow, TableCell,
  WidthType, Header, TextRun } = require("docx");
const cell = (text) => new TableCell({ children: [new Paragraph(text)] });
const doc = new Document({
  sections: [{
    headers: { default: new Header({ children: [new Paragraph("Quarterly report")] }) },
    children: [
      new Paragraph({ text: "Title", heading: HeadingLevel.HEADING_1 }),
      new Paragraph({ children: [new TextRun("Body text, "), new TextRun({ text: "bold", bold: true })] }),
      new Paragraph({ text: "A point", bullet: { level: 0 } }),
      new Table({
        width: { size: 100, type: WidthType.PERCENTAGE },
        rows: [new TableRow({ children: [cell("Day"), cell("Plan")] })],
      }),
    ],
  }],
});
Packer.toBuffer(doc).then((buffer) => fs.writeFileSync("out.docx", buffer));

Editing an existing document

docx only creates documents. To change one the user gave you, so that it keeps their template and styles, edit its XML with jszip. The body is in word/document.xml:

const fs = require("fs");
const JSZip = require("jszip");
(async () => {
  const zip = await JSZip.loadAsync(fs.readFileSync("in.docx"));
  let xml = await zip.file("word/document.xml").async("string");
  xml = xml.replace("{{CLIENT}}", "Acme Ltd");
  zip.file("word/document.xml", xml);
  fs.writeFileSync("out.docx", await zip.generateAsync({ type: "nodebuffer" }));
})();

To fill a template with placeholders, patchDocument from docx replaces {{name}} markers with paragraphs, runs or tables, and keeps everything else.

Common pitfalls

• Runs fragment. One visible sentence is often several <w:r> runs, split

mid-word wherever formatting or a spellcheck marker changed. A plain string replace over the XML misses any match that straddles runs. Find the paragraph (<w:p>) by its text, then rewrite its runs.

• Images need an explicit size. An ImageRun takes

transformation: { width, height } in pixels. Size it to the page (about 600 px wide for a portrait page with normal margins), or it inserts at the image's native size.

• Tracked changes are not applied by any of these. mammoth shows

insertions and ignores deletions. Accepting or rejecting changes means editing the w:ins and w:del elements in word/document.xml directly.

• Headers and footers belong to a section (headers/footers on it), not to

the document. In an existing file they are separate parts: word/header1.xml and the like.

• .doc (the old binary format) is not supported. Convert it with

LibreOffice first: soffice --headless --convert-to docx.