Document Text Extractor

Turn any document into clean text, Markdown, tables and metadata. Document Text Extractor reads PDF, Word, Excel, PowerPoint, EPUB, HTML, CSV and plain text files, restores the heading structure, pulls out tables, reads scanned pages with OCR, and can split everything into retrieval chunks that keep their section context. No login, no API key and no official API needed.

What is Document Text Extractor?

Document Text Extractor is a document parsing API that converts files into LLM-ready Markdown and plain text. Give it a list of document URLs, upload a file, or pass base64 data from your own code, and every document comes back as one row with the full text, the Markdown, a page-by-page breakdown, detected tables and the file's own metadata.

It exists because most document extraction tools stop at PDF and hand back a wall of unformatted text. This Actor keeps headings, lists and tables, so a language model can tell a section title from a sentence, and a retrieval system can cite the section a passage came from.

What you give it	What you get back
A PDF, Word, Excel, PowerPoint, EPUB, HTML or CSV file	Markdown, plain text, tables, chunks, metadata
A scanned PDF with no text layer	OCR text in 19 languages
A folder of mixed formats	One consistent row per document, same fields for every format

Which file formats are supported?

Format	Extensions	What is extracted
PDF	.pdf	Text per page, headings by font size, tables, metadata, OCR for scans
Word	.docx	Real heading levels, lists, tables, core properties
Excel	.xlsx, .xlsm	Every sheet as a Markdown table, sheet names
PowerPoint	.pptx	One section per slide, slide titles, tables, speaker notes
EPUB	.epub	One section per chapter, book title and author
HTML	.html, .htm	Article Markdown, page title, meta description
CSV	.csv, .tsv	Markdown table, delimiter detected automatically
Text	.txt, .md, .json	Content as-is

Password-protected PDFs are supported: pass the password in the input.

What data can Document Text Extractor extract?

Every row in the dataset describes one document.

Field	Type	Example
source	string	https://example.com/report.pdf
sourceType	string	url, upload or base64
fileName	string	report.pdf
format	string	pdf
fileSizeBytes	integer	3471
pageCount	integer	2
pagesProcessed	integer	2
title	string	Annual Operations Report 2026
author	string	Malina Data
subject	string	Sample document
creator	string	Microsoft Word
producer	string	ReportLab PDF Library
documentCreatedAt	string	D:20260904T120000Z
documentModifiedAt	string	2026-09-04T12:00:00
markdown	string	# Annual Operations Report 2026\n\n## Revenue…
text	string	Annual Operations Report 2026 Revenue grew…
wordCount	integer	198
charCount	integer	1284
pages	array	[{"number": 1, "text": "…", "markdown": "…", "ocr": false}]
tables	array	[{"page": 1, "markdown": "| Region | Q1 |…", "rows": [["Region","Q1"]]}]
tableCount	integer	1
chunks	array	[{"index": 0, "heading": "Revenue", "headingPath": ["Report","Revenue"], "text": "…", "charCount": 812}]
chunkCount	integer	4
sheetNames	array	["Sales", "Notes"]
ocrPageCount	integer	0
status	string	ok or error
errorCode	string	null, or http-error, no-text-content, too-large, …
errorMessage	string	Plain-English reason a file could not be read
warnings	array	["3 page(s) have no text layer and look scanned."]
extractedAt	string	2026-09-04T16:20:11+00:00

How much does it cost to extract text from documents?

The Actor uses pay-per-event pricing, so the bill follows the work done, not the minutes the container was alive.

Event	Price	When it is charged
Document processed	$0.0035	Once per document that produced text
Page parsed	$0.00015	Per page, sheet, slide or chapter that was read
OCR page	$0.010	Only when OCR is on and the page has no text layer

Nothing is charged for a document that failed. A dead link, a file above the size limit, a corrupt PDF or a scan with no readable text comes back as an error row at no cost. That includes the case where a scanned PDF is submitted with OCR switched off: you get a row telling you to switch OCR on, and you pay nothing for it.

Worked examples:

Job	Cost
100 born-digital PDFs, 10 pages each	$0.50
1,000 Word contracts, 3 pages each	$3.95
500 web pages to Markdown	$1.83
50 scanned PDFs, 5 pages each, OCR on	$2.71

The Apify Free plan comes with $5 of monthly usage, which is about 1,000 ten-page PDFs before you pay anything. Apify plans above the free tier get automatic discounts of 10% to 30% on every event above.

How the price compares

Comparable document Actors in Apify Store charge between $2.79 and $20 per 1,000 documents, and the two that offer OCR charge $12 and $15 per 1,000 scanned pages. This Actor is built to sit below the middle of that range while doing more per document:

For 1,000 documents	This Actor	Typical PDF-only Actor	Cheapest PDF-only Actor
3 pages each	$3.95	$5.00	$3.00
10 pages each	$5.00	$5.00	$3.00
1,000 scanned pages, OCR	$10.00	$15.00	OCR not offered
Formats handled	8	1 (PDF)	1 (PDF)
Markdown, tables, RAG chunks	all three	some	text only

The split between a document fee and a per-page fee is deliberate: a one-page invoice should not cost the same as a 300-page manual. If your documents are unusually long, the per-page part is what you are paying, and pageRange keeps it down when you only need the first pages.

There are no proxy costs: the Actor fetches files over plain HTTPS and parses them in memory, so almost none of the price is infrastructure.

How to extract text from a PDF or Word file

1. Click Try for free and sign in to Apify.
2. Paste one or more document links into Document URLs, or use Upload a file to send a file from your computer.
3. Leave Markdown output on if the text is going to an LLM; switch on Split into RAG chunks if it is going into a vector database.
4. Press Start and wait a few seconds.
5. Open the Output tab, or export the dataset as JSON, CSV or Excel.

For scanned documents, switch on OCR scanned pages and pick the language of the scan.

⬇️ Input

Only one thing is required: at least one document, given as a URL, an upload or base64 data.

{
  "urls": [
    "https://raw.githubusercontent.com/lergassy/apify-actor-assets/main/document-text-extractor/sample-report.pdf",
    "https://raw.githubusercontent.com/lergassy/apify-actor-assets/main/document-text-extractor/sample-contract.docx"
  ],
  "includeMarkdown": true,
  "includeFullText": true,
  "extractTables": true,
  "chunkForRag": false
}

Reading documents from URLs

Put direct file links in urls. The link has to point at the file itself, not at a viewer page: a Google Drive or Dropbox preview page returns HTML, not the document. Use the direct-download form of those links.

Uploading a file from your computer

Use the Upload a file field to send a single document without hosting it anywhere. This is the fastest way to test the Actor on a real file.

Passing base64 data from code or an AI agent

When your document only exists in memory, send it in base64Files. Each entry is either a plain base64 string or an object with a file name:

{
  "base64Files": [
    { "fileName": "invoice.pdf", "data": "JVBERi0xLjQKJeLjz9M…" },
    "data:application/pdf;base64,JVBERi0xLjQK…"
  ]
}

Reading only some pages

pageRange accepts 1-5, 8 or 12- (page 12 to the end). It applies to PDFs and keeps the bill down when you only need the first pages of long documents.

Splitting documents into RAG chunks

Switch on chunkForRag to get retrieval-ready chunks. Chunks break on headings and paragraphs rather than on a raw character count, tables are split by rows so every part keeps its header row, and each chunk carries the headingPath it belongs to:

{
  "index": 2,
  "heading": "Revenue by region",
  "headingPath": ["Annual Operations Report 2026", "Executive summary", "Revenue by region"],
  "text": "Revenue grew across all three regions during the year…",
  "charCount": 812
}

That path is what lets a retrieval answer say where in the document it came from, instead of returning a paragraph with no context.

Reading scanned documents with OCR

A page with no text layer is detected automatically. With ocrScannedPages off, the Actor tells you the pages are scanned and charges nothing. With it on, those pages are read by the OCR engine in the language given in ocrLanguage: English, German, French, Spanish, Portuguese, Italian, Dutch, Polish, Russian, Ukrainian, Turkish, Indonesian, Vietnamese, Arabic, Hindi, Chinese (simplified and traditional), Japanese or Korean. ocrMaxPagesPerDocument caps how many scanned pages are read per file so a 400-page scan cannot surprise you.

⬆️ Output

One row per document. Fields that were switched off in the input are simply absent.

{
  "source": "https://…/sample-report.pdf",
  "sourceType": "url",
  "fileName": "sample-report.pdf",
  "format": "pdf",
  "fileSizeBytes": 3471,
  "pageCount": 2,
  "pagesProcessed": 2,
  "title": "Annual Operations Report 2026",
  "author": "Malina Data",
  "markdown": "# Annual Operations Report 2026\n\n# Executive summary\n\nThis sample document…\n\n## Revenue by region\n\n| Region | Q1 | Q2 | Q3 | Q4 |\n| --- | --- | --- | --- | --- |\n| Europe | 120 | 140 | 155 | 162 |",
  "text": "Annual Operations Report 2026 Executive summary…",
  "wordCount": 198,
  "charCount": 1284,
  "tableCount": 1,
  "tables": [
    {
      "page": 1,
      "markdown": "| Region | Q1 | Q2 | Q3 | Q4 |\n| --- | --- | --- | --- | --- |\n| Europe | 120 | 140 | 155 | 162 |",
      "rows": [["Region", "Q1", "Q2", "Q3", "Q4"], ["Europe", "120", "140", "155", "162"]]
    }
  ],
  "ocrPageCount": 0,
  "status": "ok",
  "errorCode": null,
  "errorMessage": null,
  "warnings": [],
  "extractedAt": "2026-09-04T16:20:11+00:00"
}

Markdown that keeps the document's shape

In Word, PowerPoint and EPUB files the heading levels are already in the file, so they are used directly. PDFs carry no heading information at all, so headings are inferred from font size relative to the body text of the document, with page furniture, rotated stamps and glued-together tokens filtered out.

The same pass fixes a problem that affects academic and LaTeX-produced PDFs: when a file sets word spacing very narrow, the usual extraction returns Attentionisallyouneed instead of readable words. The Actor measures how often that happens on each page and re-reads the page with tighter word spacing when it does, so the text arrives as words rather than as one long string.

Tables

Every detected table is returned twice: as rows for code that wants arrays, and as a Markdown table for text that goes to an LLM. Columns that are empty in every row are dropped. Table detection in PDFs depends on how the table was drawn, so treat PDF tables as very good but not perfect; tables in Word, Excel, PowerPoint and CSV files are exact.

Error rows

A document that could not be read still produces a row, so a batch never silently loses a file:

errorCode	Meaning
http-error	The server refused the download or the file is gone
network-error	The host could not be reached
timeout	The download took longer than the timeout
too-large	The file is above maxFileSizeMb
unsupported-format	The file is not one of the supported formats
password-required	The PDF needs a password
corrupt-file	The file is damaged or is not what its extension claims
no-text-content	Nothing could be read, usually a scan with OCR switched off
invalid-base64	The base64 entry could not be decoded

Use cases for document data

Feeding a RAG pipeline or a knowledge base

Convert a document library into Markdown with chunks that keep their heading path, then embed the chunks. Because the heading path travels with the chunk, retrieved passages can be cited by section instead of by page number.

Contract and report review

Pull the text out of hundreds of PDFs and Word files and run a language model over the result to find dates, parties, amounts or clauses. Metadata fields tell you who produced each file and when.

Migrating a document archive

Move a folder of mixed PDF, Word and Excel files into a wiki, a CMS or a static site. Every format comes back with the same field names, so one importer handles all of them.

Reading scanned paperwork

Invoices, permits and old records that arrive as scans become searchable text with the OCR option, in nineteen languages.

Spreadsheets and data files as text

Excel and CSV files come back as Markdown tables, which is the form language models read most reliably.

Integrations

Run the Actor from the Apify Console, or connect it to the tools you already use.

• API: start a run with POST https://api.apify.com/v2/acts/lergassy~document-text-extractor/runs and read the dataset when it finishes.
• Python: ApifyClient(token).actor('lergassy/document-text-extractor').call(run_input={'urls': [...]})
• JavaScript: await client.actor('lergassy/document-text-extractor').call({ urls: [...] })
• n8n: use the Apify node, pick this Actor, and set a max total charge so a run can never exceed a budget you chose.
• Make and Zapier: run the Actor and map the dataset straight into Sheets, Airtable or Notion.
• Google Sheets: export the dataset as CSV or push it with the Sheets integration.
• MCP: the Actor is available through the Apify MCP server, so Claude, Cursor and other MCP clients can call it as a tool.
• Webhooks: trigger a webhook when a run succeeds and pass the dataset to your own service.

🤖 For AI agents and LLM apps

Call the Actor with a single field and you already get useful output:

{ "urls": ["https://example.com/report.pdf"] }

• One dataset row per document; status is ok or error, so a failed file is visible without parsing logs.
• markdown is the field to feed a model. chunks (with chunkForRag: true) is the field to embed.
• Every chunk carries headingPath, so an answer can name the section it used.
• Failed documents are never charged, which keeps agent retries cheap.
• Documents can be passed as base64 in base64Files, so an agent does not need to host a file to have it read.

Troubleshooting

The run returned http-error for a link that works in my browser. The host is refusing requests from data-centre IP addresses, or the link points at a viewer page rather than the file. Download the file and use Upload a file, or host it somewhere that allows direct downloads.

A PDF came back with no text. It is a scan. Switch on OCR scanned pages and choose the language.

The OCR text has mistakes. Check the language setting first. Low-resolution scans, handwriting and stamps are the usual limits of OCR.

My spreadsheet was cut off. Sheets are read up to 5,000 rows each, and a warning says so in warnings.

Headings are wrong in one PDF. Some PDFs set every line at the same size, and then no heading structure exists to recover. The plain text field is unaffected.

❓ FAQ

Is it legal to extract text from documents with this Actor?

The Actor only reads files you point it at. Whether you may use a given document depends on who owns it and how you obtained it, so use it on your own files, on files you are licensed to process, and on publicly available documents. It is not a substitute for legal advice.

Do I need an API key or a login?

No. There is no account, key or cookie to configure. You need an Apify account to run any Actor, and nothing else.

How much does it cost to extract text from 1,000 PDFs?

For 1,000 PDFs of ten pages each, $3.50 for the documents plus $1.50 for the pages, so $5.00 in total. Scanned pages cost $0.01 each on top, and only when OCR is on.

Can I use it with the Apify API?

Yes. Start a run with the API, then read the dataset. The Actor is a normal Apify Actor with a standard input schema.

Can I use it through an MCP server?

Yes. It is exposed through the Apify MCP server, so an MCP client such as Claude or Cursor can call it directly as a tool.

Can I integrate it with other apps?

Yes: n8n, Make, Zapier, Google Sheets, Slack, Airtable and anything that can call an HTTP API or receive a webhook.

Does one row mean one document or one page?

One row is one document. Pages live inside that row in the pages array when Per-page output is on, and chunks live in chunks.

What is the largest file it can read?

The default limit is 50 MB per file, adjustable up to 200 MB. Files above the limit are reported as too-large and are not charged.

Can it read password-protected PDFs?

Yes, if you supply the password in the PDF password field.

Does it work with Google Docs links?

Not with a share link, because that returns a web page. Use the Download as PDF or Word form of the link, or upload the exported file.

Your feedback

Found a document the Actor reads badly, or a format you need? Open an issue on the Issues tab or leave a review. Issues are answered quickly, and format requests are the main way this Actor grows.

You might also like

Actor	What it does
Email & Phone Verifier	Validate e-mail addresses and phone numbers, find contacts on a website
US Business Leads	New business registrations from US state and city registries
Google Flights Scraper	Flight prices, airlines, layovers and booking links
Agoda Reviews Scraper	Hotel reviews and ratings from Agoda and Booking.com