Pi Create Document

PDF, Word, PowerPoint, Excel, CSV and PNG files from Pi, with nothing to install.

Agents write text and Markdown well, but a report as a PDF, a deck or a spreadsheet usually means LibreOffice, Pandoc, a headless browser or Python packages, installed on the spot. This extension gives the agent three tools that create, convert and read those formats in-process. Fonts, diagram rendering and every file format are bundled; nothing is downloaded and no system tool is run.

• create_document writes a file from Markdown (or HTML, text, CSV, TSV, JSON, Mermaid or SVG). The extension picks the format: report.pdf, report.docx, deck.pptx, data.xlsx, flow.png.
• convert_document converts a file between formats. PDFs and documents become one PNG per page, so the agent can look at what it made.
• read_document reads PDF, Word, PowerPoint, Excel and other files as Markdown, so the agent can check its output or work from a file it was given.

Install

From this checkout, after installing dependencies at the repository root:

pi install /path/to/pi-extensions/pi-create-document

Then restart Pi or run /reload. For a one-session preview:

pi -e /path/to/pi-extensions/pi-create-document/index.ts

Published releases:

pi install npm:@nothumanwork/pi-create-document

Requires Pi 1.4.0 and Node.js 22.19+. Two dependencies are native: @napi-rs/canvas (PDF pages to PNG) and @resvg/resvg-js (SVG to PNG). npm installs their prebuilt binaries for Linux, macOS and Windows on x64 and arm64. Neither runs an install script.

Formats

Word, PowerPoint and Excel output is standard Office Open XML, the format Microsoft Office uses and Google Docs, Slides and Sheets import; it was checked by rendering it in LibreOffice. Workbooks ask the application to work out their formulas on opening.

Writing documents

Content is GitHub-flavoured Markdown:

• Headings, emphasis, strikethrough, inline code, links, lists, task lists, tables with column alignment, block quotes, code blocks, footnotes, horizontal rules and images (local paths or data: URIs) all carry over.
• YAML front matter sets the metadata: title, author, subject, description, keywords and lang.
• A ```mermaid block becomes a diagram: flowchart/graph, sequenceDiagram, classDiagram, stateDiagram, erDiagram, xychart and pie. Other Mermaid types stay as code, with a warning. A ```svg block becomes a figure.
• <!-- pagebreak --> (or \newpage) starts a new page.
• A single level 1 heading at the top is the title: Word's Title style, the PDF and HTML title, the deck's cover slide. With toc: true, PDF, Word and HTML get a table of contents after it. In Word it is a real table-of-contents field, written with its page numbers.

Slides. A new slide starts at each ---, page break, and heading at the slide level: H2 under a single H1 (which becomes the cover), otherwise the top heading level used. Content that does not fit continues on further slides titled "(continued)". A short text beside one picture is laid out in two columns. HTML comments and footnotes become speaker notes.

Spreadsheets. Each Markdown or HTML table becomes a sheet named after the heading above it, or give CSV, TSV or JSON. Cells become typed values: numbers with thousands separators, currency and percentages keep their display format, ISO dates become dates, TRUE/FALSE become booleans and =SUM(B2:B9) becomes a formula. Identifiers with leading zeros, phone numbers and version strings stay text.

Images. Relative paths resolve against the session directory for create_document, or the input's folder for convert_document. Remote http(s) images are never fetched; download them first. A skipped image is a warning, not a failure.

Tool options

Several PNG pages are written as name-1.png, name-2.png, … numbered by page. Writes go through Pi's file mutation queue and a temporary file, so a failed conversion leaves nothing half-written. A conversion never replaces its own input.

Limits

• Legacy .doc. No JavaScript library writes Word 97–2003 binary files. A .doc output holds Rich Text, which Word and LibreOffice open as a Word document; the result says so. Write .docx for full fidelity and for Google Docs. Reading a real .doc keeps only its text. Excel and PowerPoint 97–2003 files (.xls, .ppt) cannot be read; save them as .xlsx or .pptx.
• Reading PDFs recovers text, headings and paragraphs, and drops running headers, footers and contents pages. Tables, columns and images are not rebuilt. To see a PDF, convert it to PNG.
• Fonts. Text is set in Roboto and code in Roboto Mono, with Noto fallbacks for maths, symbols and emoji. Emoji are monochrome in PDF and PNG output. Word, RTF and PowerPoint output names Calibri and Consolas, which the viewing application supplies.
• Word contents page numbers come from laying the pages out as Word does. When the layout meets something it cannot measure exactly, such as emoji, it estimates the rest and the result warns. Word corrects the numbers when the table is updated.
• Inputs are limited to 200 MB. PNG output is limited to 10,000 pixels a side.

SDK

The conversion code works without Pi:

import { convertDocument, createDocument, readDocument } from "@nothumanwork/pi-create-document";

await createDocument({ content: "# Q3\n\n| Region | Revenue |\n|---|---:|\n| EMEA | $1,250 |\n", output: "q3.xlsx" });
await convertDocument({ input: "q3.xlsx", output: "q3.pdf" });
const { markdown } = await readDocument({ path: "q3.pdf" });

renderMermaid, svgToPng, pdfToPng and readPdf are exported too. The Pi extension is the ./pi export.

Dependencies

Documents are written with pdfmake, docx, PptxGenJS and ExcelJS. They are read with PDF.js, mammoth and word-extractor. Diagrams come from beautiful-mermaid, and Markdown is handled by the unified mdast and hast utilities. RTF reading and writing, and pie charts, are this package's own. The dependencies are under permissive licences (MIT, ISC, BSD, Apache-2.0 and similar), except two with file-level copyleft: @resvg/resvg-js (MPL-2.0) and elkjs (EPL-2.0), which beautiful-mermaid uses for layout. The bundled fonts are under the SIL Open Font License.

npm audit reports two advisories in this package's dependency tree, neither reachable from it:

• image-size (GHSA-5p2g-fcmc-qvqq, GHSA-w3rx-r6r6-pgpr) is declared by PptxGenJS 4.0.1 but not imported by either of its builds.
• uuid before 11.1.1 (GHSA-w5hq-g745-h8pq) affects v3, v5 and v6 called with a buffer. ExcelJS 4.4.0 only calls v4 without one.

npm audit fix --force would downgrade both libraries to releases with breaking changes, so it is not applied.

Development

npm run check --workspace @nothumanwork/pi-create-document
npm test --workspace @nothumanwork/pi-create-document

The tests convert a sample report to every format and read each back. They stub fetch to show that nothing reaches the network.