Office Documents to Markdown – DOCX, PPTX, XLSX (and ODT, ODP, ODS) for LLMs and RAG

Give it links to Word, PowerPoint and Excel files and get clean Markdown, plain text, every table as JSON, speaker notes and metadata back, one record per file. No browser, no OCR, no API key, no LLM: the files are read directly, so results are fast and the same file always gives the same answer.

Price: $0.0015 per converted document (that is $1.50 per 1,000 documents), one flat price per file whether it has 2 pages or 200 slides. Files that cannot be converted are free.

Who it is for

• RAG and AI-agent builders who need Office files as text a model can read, with headings, lists and tables still in place and metadata to cite.
• Teams that index documents: intranets, knowledge bases, search, compliance archives, public-sector data portals.
• Data people who want the tables of Word files and the sheets of workbooks as JSON, not as a screenshot.
• Scheduled jobs that watch a list of public documents (the sha256 of each file tells you when one changed).

For PDFs use the sister Actor PDF to Text & Markdown Extractor (far003/pdf-to-text-markdown-extractor): same record style, same price idea. If you send a PDF here you get a free record that says so.

What it reads

File	What you get
Word (.docx, .docm, .dotx)	Headings (from styles and outline levels), paragraphs, bold / italic / strike / super- and subscript, links, bullet and numbered lists (nested, with Word's own numbering, "2.1" legal numbers kept), tables with merged cells, footnotes and endnotes, text boxes, code in monospace fonts, quotes. Tracked changes are read as accepted (deletions out).
PowerPoint (.pptx, .ppsx, .pptm)	One section per slide, in presentation order: the title, the text in reading order (top to bottom, then left to right), bullets with their levels, tables, charts as a table of their values, SmartArt text, alt text of pictures, and the speaker notes. Hidden slides are kept and marked.
Excel (.xlsx, .xlsm, .xltx)	One table per sheet. Cells show what Excel shows: dates as dates, 12.5% as 12.5%, 1234.5 as 1,234.50 where the cell format says so. Formulas use the value Excel saved. Empty rows and empty leading columns are dropped. The row limit is yours to set and the record says when it cut.
OpenDocument (.odt, .odp, .ods)	The same output as the three above. Spreadsheet cells carry their displayed text.

The type is recognised from the content of the file (the ZIP package and its content types), never from the extension or the Content-Type the server sends: a DOCX served as report.pdf is still read as a DOCX.

Quick start

1. Paste the links into Document URLs.
2. Click Start.
3. Open the Overview tab for the list, Markdown and text for the content, Tables (JSON), Slides and notes, or download the dataset as JSON, CSV or Excel.

Example input

This is the real example prefilled in the input form:

{
  "urls": [
    "https://www.epa.gov/system/files/documents/2025-05/drinking-water-erp-template.docx",
    "https://raw.githubusercontent.com/apache/poi/trunk/test-data/slideshow/SampleShow.pptx",
    "https://www.irs.gov/pub/irs-soi/23dbalabama.xlsx"
  ]
}

Run result: 3 records (3 ok), 3 charged = 3 × $0.0015 = $0.0045.

Example output (a real run, shortened)

A Word file (a US EPA template, 29 pages, 38 tables). Every field is always present; fields that do not apply are null:

{
  "url": "https://www.epa.gov/system/files/documents/2025-05/drinking-water-erp-template.docx",
  "status": "ok",
  "error": null,
  "charged": true,
  "httpStatus": 200,
  "fileName": "drinking-water-erp-template.docx",
  "sizeBytes": 217904,
  "sha256": "7f5e54a224a47ffa81cf1269996d7847c531cf1051d5fc1dc9510d12b95f7ea3",
  "format": "docx",
  "title": "Emergency Response Plan Template",
  "titleSource": "metadata",
  "language": "en-US",
  "createdAt": "2025-05-19T13:39:00.000Z",
  "modifiedAt": "2025-05-19T17:27:00.000Z",
  "application": "Microsoft Office Word",
  "pageCount": 29,
  "tableCount": 38,
  "imageCount": 0,
  "wordCount": 5407,
  "hasMacros": false,
  "truncated": false,
  "warnings": []
}

Its markdown field (an excerpt of the real output):

### Utility Information

| **PWSID** | *123456* |
| --- | --- |
| **Owner** | *Town of Anytown* |
| **Total population served and total service connections** | *7,500 served - 2,500 connections* |
☐ Site plans and “as built” drawings for the following components of your system (as applicable):

    - Pumping and storage facilities
    - Reservoir facilities
    - Water treatment facilities

A PowerPoint file (Apache POI sample deck): markdown has a section per slide, notes included, and slides[1] is

{
  "number": 2,
  "title": "This is the second slide",
  "hidden": false,
  "text": "- It has bullet points on it\n- They’re fun, aren’t they?\n- Especially in a different font like Arial Black at 16 point!",
  "notes": "These are the notes of the 2nd slide\nTHIS LINE IS BOLD"
}

An Excel file (IRS statistics): markdown is a table, tables[0].rows the same cells as arrays, sheets the facts about the sheet:

{ "name": "Sheet1", "state": "visible", "rowCount": 17, "columnCount": 6, "rowsReturned": 17, "truncated": false }

## Sheet: Sheet1

| Alabama, Fiscal Year 2023 |  |  |  |  |  |
| --- | --- | --- | --- | --- | --- |
|  | Number of returns filed | Number of returns filed electronically | Internal Revenue gross collections (thousands of dollars) | ... |
| Total [1] | 3,471,529 | 2,690,994 | 35,829,649 | 1,706,757 | 6,809,200 |
| Estate tax | 391 | n/a | 309,381 | 27 | 12,315 |

Input

Field	Default	What it does
urls	–	Addresses of the files. Bad entries get a free invalid record, they never stop the run.
outputFormats	all four	markdown, text, tables (every table as JSON), slides (one object per slide). Choose fewer for lighter records.
maxRowsPerSheet	1000	Spreadsheets: non-empty rows converted per sheet (up to 100,000). The record tells you the total and that it was cut. The memory of the run can cut earlier on a very big sheet (see the FAQ).
includeHiddenSheets	off	Hidden sheets are skipped (and the warnings say so) unless you turn this on.
includeSpeakerNotes	on	Notes after each slide.
maxFileMegabytes	50	Bigger files are refused (free, too_large), the download is stopped as soon as the limit is passed.
respectRobotsTxt	on	The robots.txt of the site is read before the first request and followed for every redirect.
requestsPerSecond	2	Per host. A Crawl-delay in robots.txt can only slow it down.
maxConcurrency	3	Documents converted at the same time.
timeoutSecs	60	Download plus conversion, per document.

Output fields

One record per input address, in the Overview view; the same record has everything.

Field	Meaning
url, finalUrl	The address as given and after redirects (credentials removed).
status	See the table below.
error	Why there is no (or only a partial) result, in one sentence; null when all went well.
charged	true only for ok and partial.
httpStatus, fileName, sizeBytes, sha256	About the download. sha256 is the hash of the file: use it to detect changes.
format	docx, pptx, xlsx, odt, odp or ods, as found in the content.
title, titleSource	The title of the file properties; if there is none, or it is a program default ("Document1", "Microsoft Word - …"), the first heading (first-heading) or first slide title (first-slide-title).
subject, keywords, language	From the file properties; language also from the language of the text when the file does not say (BCP 47, e.g. en-US).
createdAt, modifiedAt, application	Dates (ISO 8601, UTC) and the program that made the file.
pageCount	Word and Writer only, as declared by the file (Word does not store the layout's pages, so the number may be stale); null otherwise.
slideCount, sheetCount	Slides; sheets (hidden ones included).
tableCount, imageCount	Tables converted (one per worksheet); pictures found (not converted).
wordCount, charCount	Of the converted text.
hasMacros	The file holds macros. They are never read or run.
truncated, fullTextKey	true when rows were cut or the text was too long for a record; then the whole Markdown is in the key-value store under fullTextKey.
markdown, text	The document as Markdown and as plain text.
tables	[{ index, name, location, rowCount, columnCount, rows }], rows being arrays of cell texts.
slides	PPTX/ODP: [{ number, title, hidden, text, markdown, notes }].
sheets	XLSX/ODS: [{ name, state, rowCount, columnCount, rowsReturned, truncated }].
warnings	What the conversion could not carry over (see below).
extractedAt	When the record was made.

Statuses

status	Meaning	Charged
ok	Converted.	yes
partial	Converted, but a part could not be read (a damaged slide or sheet; the warnings name it).	yes
empty	A valid file with no text (only pictures, or blank).	no
password_protected	The file is encrypted. It cannot be read.	no
unsupported_format	Legacy .doc / .xls / .ppt, RTF, EPUB, XLSB, OpenDocument drawings… The message says what it is and to save it as DOCX / PPTX / XLSX.	no
not_office	The URL is not an Office document: an HTML page, a PDF, a ZIP, an empty answer.	no
corrupt	A damaged or truncated file.	no
unsafe_archive	Refused for safety: a zip bomb (absurd compression ratio, or parts larger than they declare).	no
too_large	Over maxFileMegabytes, or a document whose XML is too big to be read safely.	no
blocked	robots.txt forbids it, the site answered 401/403/429 or a bot-challenge page, or the address leads to a private network. Nothing is done to get around it.	no
unreachable	The download failed (DNS, timeout, 404, 5xx after one retry).	no
invalid	Not a usable web address.	no

Warnings: what is not carried over

The conversion tells you, record by record, what it left out, so you do not have to guess:

• Pictures (counted in imageCount; the author's alt text is kept when it says something) – there is no OCR and no image description.
• Charts in Word and Excel (in PowerPoint a chart becomes a table of its values), equations (read as plain text), SmartArt (its text only), comments, headers and footers, the table of contents (the headings are there).
• Formulas are not calculated: the value Excel saved is shown; a formula without a saved value is an empty cell, and the warning counts them.
• Macros are flagged (hasMacros) and never read.
• Hidden sheets, rows cut at maxRowsPerSheet (or by the memory of the run on a very big sheet), columns beyond 500.

Pricing

One event: document – $0.0015 for each file converted (ok or partial). Nothing else is charged: no start fee, no per-page, per-slide or per-row price.

You convert	Cost
1 document	$0.0015
100 documents	$0.15
1,000 documents	$1.50
1,000 addresses where 150 are broken links or protected files	$1.275 (850 × $0.0015)

Platform usage (compute) is included in that price: you pay nothing else.

Spending limit: if you set a maximum cost per run, the Actor stops cleanly when it is reached ("spending limit reached" in the run summary) and does not download more files.

Safe and polite by design

• Privacy: the author and last editor fields of the file properties are read by nobody: they are not in the output (personal data).
• robots.txt is respected (checked for every site and every redirect), with an honest User-Agent that names the bot. No proxies, no disguise, no CAPTCHA tricks: a block page is reported as blocked.
• Private networks are never reached, not even through a redirect or a name that resolves to a private address.
• Zip bombs and absurd files are refused: the size is checked while the file is unpacked, not just as the header declares it.
• Nothing is executed. Macros, scripts, links and embedded objects are never run or followed.
• A restarted run does not deliver or charge the same document twice.

FAQ

Does it do OCR? No. A scanned image inside a Word file stays an image (counted, with its alt text if any). For text in images use an OCR Actor.

Can it read an old .doc, .xls or .ppt? No, those are a different (binary) format. You get a free record saying so; save the file as DOCX / XLSX / PPTX once and run again.

What about password-protected files? They cannot be read; the record is password_protected and free.

Why is pageCount null for my Word file? Word stores the page count only when it saved the file with it. We report what the file declares and never invent a number. Slides and sheets are always counted.

Are the numbers in my spreadsheet what Excel shows? Dates, percentages, thousands separators, decimals and currency follow the cell format. Formulas are not recalculated. If a number looks different from what you see in Excel, tell us the format code.

Is there a memory limit? Yes, and a big file never crashes the run. Spreadsheets (XLSX, ODS) are read as a stream, so the size of the file is not the limit: the rows you get back are, because they are kept in memory until the record is written. With the default 1 GB that is about 40,000 rows of 10 short columns (numbers, dates) or about 18,000 rows of 17 text columns (a long public register is that kind of file). Past that the sheet is cut: the record says truncated: true, sheets[].rowCount has the real total, and a warning names the sheet ("Rows cut to stay within the memory of the run: give the run more memory or lower maxRowsPerSheet"). A cut file is still converted and charged like any other; the other files of the run are not affected. More memory gives proportionally more rows (4 GB about four times as many). Files refused for safety are free (too_large) with the reason in error: an XLSX or ODS sheet over 256 MB unpacked, an XLSX whose table of shared strings does not fit in a quarter of the run's heap (about 90 million characters of text on 1 GB), a Word, PowerPoint, Writer or Impress file whose XML is over 30 MB unpacked, or a package over 400 MB unpacked.

My workbook is huge. The first maxRowsPerSheet non-empty rows of each sheet are converted, and sheets[] has the real row counts. Raise the limit (up to 100,000) if you need more, and give the run more memory if the record says the rows were cut by memory; very long outputs are saved in the key-value store.

Does it follow links inside the documents? No. Links are kept as Markdown links; only the addresses you give are downloaded.

Is a file I put on a private server readable? Only public http(s) addresses. Private and local addresses are refused by design, and credentials in a URL are removed and not used.

How does it compare with Microsoft's MarkItDown? MarkItDown is an excellent open-source library; this Actor is the hosted version for batches: URLs in, records out, with the safety limits, the per-file status, the tables as JSON, the notes per slide and the "what was lost" warnings built in.

Integration

Run it from the Apify API, the CLI, or your own code.

JavaScript

import { ApifyClient } from 'apify-client';

const client = new ApifyClient({ token: process.env.APIFY_TOKEN });
const run = await client.actor('far003/office-docs-markdown').call({
    urls: ['https://example.org/handbook.docx', 'https://example.org/q3-results.pptx'],
    outputFormats: ['markdown', 'tables'],
});
const { items } = await client.dataset(run.defaultDatasetId).listItems();
for (const doc of items) {
    if (doc.status === 'ok') console.log(doc.title, doc.wordCount, doc.markdown.slice(0, 200));
    else console.warn(doc.url, doc.status, doc.error);
}

Python

from apify_client import ApifyClient

client = ApifyClient("<APIFY_TOKEN>")
run = client.actor("far003/office-docs-markdown").call(run_input={
    "urls": ["https://example.org/handbook.docx"],
    "maxRowsPerSheet": 5000,
})
for doc in client.dataset(run["defaultDatasetId"]).iterate_items():
    print(doc["status"], doc["format"], doc["title"], doc["wordCount"])

Chunking for RAG: split markdown on \n## for slides and sheets, on headings for Word files; slides[].markdown and tables[] are already cut at natural boundaries.

Run summary (totals per status, charged, skipped) is stored under the key SUMMARY of the default key-value store.

Related

• PDF to Text & Markdown Extractor – the same job for PDFs: reading order, columns, headings, tables, bookmarks.