PDF & Scanned Document to Text, Tables & Fields (OCR)

Turn PDFs, scanned documents and images into clean text, LLM-ready Markdown, tables (rows and columns) and key fields such as dates, money amounts, percentages and invoice or order numbers. Paste a list of document URLs and get one structured JSON record per document, ready to export as JSON, CSV or Excel or to pass to your RAG pipeline, spreadsheet or database.

• Text PDFs: reads the PDF's own text layer (fast and exact).
• Scanned PDFs and images: built-in OCR (Tesseract) in 11 languages. In auto mode, OCR runs only on pages that need it.
• Tables: detects ruled tables and returns them as a header plus rows, and as Markdown tables.
• Key fields: dates normalized to ISO 8601, amounts with currency, percentages, URLs, reference numbers, and Label: value pairs.
• No external AI, no API key, no account needed. Your documents are processed inside the Actor run and are not sent to any third-party service.

What can I use it for?

• RAG and AI agents: convert PDF reports, manuals and filings into Markdown chunks with tables kept intact.
• Invoice and receipt pre-processing: pull totals, dates and invoice numbers from batches of PDFs or scans.
• Statistics and government reports: get the tables from PDF releases (census, central bank, ministry reports) as rows you can load into a spreadsheet.
• Digitizing scans: make scanned letters, forms, book pages and faxes searchable.

How to use it

1. Add one or more direct links to PDF or image files (.pdf, .png, .jpg, .tif, .webp, .gif, .bmp). Up to 1,000 per run.
2. Pick the OCR languages used in your scans (for example English + German).
3. Click Start. Each document becomes one row in the dataset.

Example input

{
    "documentUrls": [
        { "url": "https://www.irs.gov/pub/irs-pdf/fw9.pdf" },
        { "url": "https://raw.githubusercontent.com/tesseract-ocr/tessdoc/main/images/eurotext.png" }
    ],
    "ocrMode": "auto",
    "ocrLanguages": ["eng", "deu"],
    "extractTables": true,
    "extractFields": true,
    "pageRange": "",
    "maxPagesPerDocument": 50
}

Input options

Field	Default	What it does
documentUrls	required	Direct public links to PDF or image files.
ocrMode	auto	auto = OCR only pages without a usable text layer. always = OCR every page (for PDFs with a broken text layer). never = text layer only.
ocrLanguages	["eng"]	Up to 4 of: English, German, French, Spanish, Italian, Portuguese, Dutch, Polish, Czech, Slovenian, Croatian.
extractTables	true	Return tables from text-based PDF pages.
extractFields	true	Return dates, amounts, percentages, URLs, reference numbers and label/value pairs.
includeFullText	true	Add the whole document as text and markdown.
includePages	true	Add per-page text, Markdown, method and OCR confidence.
pageRange	all pages	For example 1-3, 5, 10-.
maxPagesPerDocument	50	Safety limit per document (max 1,000).
maxFileSizeMb	50	Larger files are skipped with a clear error.
maxConcurrency	2	Documents processed in parallel.

Output

One dataset item per document. The example below is a real result for the U.S. Census Bureau's New Residential Construction release (shortened with ... and to the first item of each list). Every item has the same fields, including failed ones (status: "error" with a plain-language errorMessage), so your exports always have the same columns.

{
    "url": "https://www.census.gov/construction/nrc/pdf/newresconst.pdf",
    "fileName": "newresconst.pdf",
    "status": "ok",
    "errorMessage": null,
    "fileType": "pdf",
    "fileSizeBytes": 329096,
    "pageCount": 7,
    "pagesProcessed": 7,
    "truncated": false,
    "ocrPageCount": 0,
    "tableCount": 10,
    "metadata": { "title": "Microsoft Word - newresconst_auto_text.docx", "subject": null, "keywords": null, "creatorTool": "PScript5.dll Version 5.2.2", "producer": "Acrobat Distiller 26.0 (Windows)", "createdAt": "2026-10-15T08:53:41", "modifiedAt": "2026-10-15T11:28:35" },
    "text": "FOR RELEASE AT 8:30 AM EDT, TUESDAY, OCTOBER 20, 2026\nMONTHLY NEW RE ...",
    "markdown": "FOR RELEASE AT 8:30 AM EDT, TUESDAY, OCTOBER 20, 2026\nMONTHLY NEW RE ...",
    "fields": {
        "dates": [
            { "raw": "OCTOBER 20, 2026", "iso": "2026-10-20", "ambiguous": false },
            { "raw": "October 20, 2026", "iso": "2026-10-20", "ambiguous": false }
        ],
        "amounts": [],
        "percentages": [
            { "raw": "76.3%", "value": 76.3 },
            { "raw": "-0.2%", "value": -0.2 }
        ],
        "urls": ["www.census.gov/economic-indicators/", "www.census.gov/construction/nrc/"],
        "references": [],
        "keyValues": [
            { "label": "Release Number", "value": "CB26-162" },
            { "label": "Building Permits", "value": "1,400,000 1,500" }
        ]
    },
    "tables": [
        {
            "pageNumber": 3,
            "tableIndex": 1,
            "title": null,
            "rowCount": 21,
            "columnCount": 12,
            "header": ["", "", "United", "States", "", "", "Northeast", "", "Midwest", "", "South", "West"],
            "rows": [
                ["September", "1,444", "888", "58", "498", "148", "57", "204", "130", "773", "517", "319"],
                ["October", "1,418", "886", "55", "477", "145", "58", "197", "123", "736", "516", "340"]
            ],
            "markdown": "|  |  | United | States |  |  | Northeas ..."
        }
    ],
    "pages": [
        { "pageNumber": 1, "method": "text-layer", "ocrConfidence": null, "charCount": 2544, "tableCount": 0, "text": "FOR RELEASE AT 8:30 AM EDT, TU ...", "markdown": "FOR RELEASE AT 8:30 AM EDT, TU ...", "processingMs": 92 }
    ],
    "warnings": [],
    "processingMs": 1900
}

• pages[].method is text-layer or ocr. ocrConfidence is Tesseract's mean word confidence (0–100) for OCR pages, so you can flag poor scans.
• fields.dates[].ambiguous is true for dates like 03/04/2024, which could be day-first or month-first. These are read as day-first.
• The dataset has two views in the Console: Overview (status, pages, tables, errors) and Text & fields.

Pricing (pay per event)

You pay only for what is extracted. Failed downloads and unreadable files are not charged.

Event	Price
Document processed	$0.001 per document
Text page (text-layer PDF page, incl. tables and fields)	$0.0003 per page
OCR page (scanned page or image)	$0.004 per page

Examples:

• 1,000 single-page text PDFs: about $1.30.
• 100 ten-page reports: $0.40.
• 1,000 scanned single-page images: $5.00 (100 ten-page scanned PDFs: $4.10).

Platform usage is included in these prices. Set a maximum cost per run in the run options and the Actor stops cleanly when it is reached; documents near the limit are cut to the pages you can still afford, with a warning on the item.

FAQ

Does it work on scanned PDFs? Yes. In auto mode, every page with no usable text layer is rendered at 300 DPI and read with OCR. Mixed PDFs (some scanned pages, some digital pages) are handled page by page.

Does it use ChatGPT or another LLM? No. Extraction is deterministic: a PDF parser, a table detector, Tesseract OCR and rule-based field extraction. The output is repeatable and nothing is sent to an AI provider. The Markdown output is designed to be fed to your own LLM.

Can it extract tables from scanned documents? Not yet. Tables are detected on text-based PDF pages that have ruled lines or cell borders. On scanned pages you get the text in reading order.

Which links work? Links that download the file directly, such as most .pdf URLs, raw GitHub files and S3 or CDN links. Viewer pages (Google Drive preview, DocSend, websites that embed a PDF viewer) return HTML; the Actor reports this clearly instead of failing silently. Files behind a login are not supported.

Is personal data extracted? The field extractor deliberately skips e-mail addresses, phone numbers and labels such as name, address, birth date, IBAN or signature, and the PDF Author metadata is omitted. The full text still contains whatever the document contains, so only process documents you are allowed to process.

Password-protected PDFs? They are reported as an error ("password-protected") and not charged.

How fast is it? Text PDFs take about 0.1–0.6 s per page. OCR takes about 2–8 s per page at the default 1 GB memory. For large OCR batches, use 2–4 GB memory and raise maxConcurrency.

Limits

• Max 1,000 URLs per run, 1,000 pages per document, 200 MB per file.
• OCR quality depends on the scan. Low-resolution photos, handwriting and heavy skew reduce accuracy; check ocrConfidence.
• Table detection needs visible ruled lines or cell borders. Tables laid out only with whitespace come through as text.
• When only the number columns of a table are ruled (common in statistical releases), the row labels to the left are added as the first column. Multi-line column headers come out as several header rows.
• Field extraction is rule-based: numeric dates, amounts and percentages work in any language, but written month names are recognized in English only. Reference-number labels cover EN/DE/FR/ES/IT/SL. It won't find every field on every layout.