PDF Tables to JSON with Cell Coordinates + RAG Chunks (OCR)

Give this Actor PDF URLs and get structured JSON back. Two things set it apart from the usual PDF text extractor:

1. Tables as cells with coordinates. Every table comes back as rows (an array of rows, each row an array of cell strings, already split), with its page number, a bbox [x0, top, x1, bottom] in PDF points, a detected header, a GitHub Markdown rendering, and, on request, a bbox for every single cell. You can link any value back to the exact spot on the page.
2. RAG-ready chunks with page references. Page text is split into sentence-aligned chunks of a size you choose, each carrying page, chunkIndex, charStart, charEnd and a citation object, so citations in a retrieval pipeline point at a page and a character range, not just a file. The chunks use the same schema as the sibling Actor transcribe-to-json-chunks, which returns recordings as chunks with time citations, so one pipeline can ingest PDFs and audio or video the same way.

On top of that you get each page's plain text, a Markdown version (headings by font size, tables in place), PDF metadata, and optional OCR for scanned pages. It is a PDF to JSON converter, a PDF table extraction tool, a PDF parser, a PDF to Markdown converter and a RAG chunker in one call, priced per page.

What it does

• Downloads each PDF from the URL you give (up to 100 per run, 50 MB each by default), checks that it really is a PDF, and parses it page by page. Nothing is stored after the run: the file is deleted as soon as it is parsed.
• Tables: ruled tables are found by their lines; when ruled columns have no row lines the cells are re-read by text rows; tables without any lines are found by text alignment with strict quality filters (at least 3 rows, 2 columns, half the cells filled, some numeric content) so paragraphs do not come back as fake tables. Overlapping detections are merged, tables under 2 rows or 2 columns are dropped. Each table has tableIndex, page, bbox, rowCount, columnCount, header (array or null), rows, markdown, strategy, and with includeCellBoxes a cells array of {row, col, text, bbox}.
• Text: reading-order text per page with line breaks. Pages whose generator wrote no space glyphs (common with TeX and some report tools) still come back with proper word spacing.
• Markdown: lines whose font is 20 percent above the page's body text become ## headings, wrapped lines are joined into paragraphs, tables are rendered as Markdown tables where they appear.
• Chunks: chunkSize characters (default 1200, about 300 tokens) with chunkOverlap (default 150), cut at sentence ends when possible. Offsets are into that page's text. Each chunk carries citation: {"type": "page", "page": 2, "label": "p. 2"} (see Family below).
• OCR: off (default), auto (only pages with fewer than 20 characters of text layer) or always. Pages are rendered at 200 dpi and read with Tesseract. OCR pages return text, Markdown paragraphs and chunks; with includeWordBoxes also word boxes with confidence. About 2 to 3 seconds per page.
• Two output shapes: pages (one dataset item per page, the default, ideal for RAG) or documents (one item per PDF with metadata, all text joined by form feeds, and every table and chunk tagged with its page).
• A PDF that cannot be fetched or read never fails the run: it produces one item with error and errorType (http_403, http_404, not_pdf, too_large, timeout, encrypted, ...) and is not charged. The other PDFs still return data.

Who it is for

• AI agents and RAG pipelines that need page-cited chunks and real tables instead of a wall of text. Call it through the Apify MCP server (see below) and index the chunks array as is.
• Data teams turning recurring PDF reports into rows: statistical releases, price lists, schedules, invoices with a text layer.
• Finance and research analysts pulling tables out of filings, central bank releases and research reports, with coordinates to verify every number against the page.
• Compliance and legal teams that must quote a page and position, not just a document.
• Anyone who needs PDF to Markdown or OCR PDF to text for a handful of pages without installing anything.

Quick start

1. Open the Actor in Apify Console and press Start. The default input parses a public Federal Reserve statistical release (12 pages, several tables) in a few seconds.
2. Replace the URL with your own PDF link. Use pages like "1-5,8" to parse only what you need.
3. For scanned PDFs set ocr to auto (mixed documents) or always (fully scanned).
4. For one JSON object per file set outputMode to documents.
5. Read results from the default dataset (JSON, CSV or Excel) or the API: https://api.apify.com/v2/datasets/<datasetId>/items. The SUMMARY record in the key-value store lists every file's status, pages processed and charged, OCR pages and errors.

Input example

{
    "pdfUrls": ["https://www.federalreserve.gov/releases/h41/current/h41.pdf"],
    "pages": "1-3",
    "outputMode": "pages",
    "extractTables": true,
    "markdown": true,
    "chunkSize": 1200,
    "chunkOverlap": 150,
    "ocr": "off",
    "includeCellBoxes": false
}

All fields are optional. A run with no input parses the Federal Reserve release above.

Field	Type	Default	Meaning
pdfUrls	array of strings	one sample PDF	Public http(s) links to PDF files, up to 100
pdfUrl	string	empty	Shortcut for a single URL, merged into pdfUrls
pages	string	all	1-based pages like "1-5,8" or "3-"
maxPagesPerPdf	integer 1 to 1000	50	Cap on pages processed and charged per file
outputMode	pages or documents	pages	One item per page, or one per PDF
extractTables, extractText, markdown	booleans	true	Turn parts of the output off
chunkSize	integer	1200	Chunk size in characters, 0 disables chunks
chunkOverlap	integer	150	Characters shared by consecutive chunks
ocr	off, auto, always	off	When to run OCR
ocrLanguage	string	eng	Tesseract language, only English is installed
maxFileSizeMb	integer 1 to 200	50	Larger files are skipped with an error item
includeWordBoxes	boolean	false	Add every word with its box to each page
includeCellBoxes	boolean	false	Add a box for every table cell

Output example

One page item from the default run (page 2 of the Federal Reserve H.4.1 release; text, markdown, rows and chunks are shortened here):

{
    "sourceUrl": "https://www.federalreserve.gov/releases/h41/current/h41.pdf",
    "documentId": "6cb2cd9838d643a55fd6591794268e74e7b0224d",
    "fileName": "h41.pdf",
    "page": 2,
    "pageCount": 12,
    "width": 612,
    "height": 792,
    "text": "FEDERAL RESERVE statistical release\nH.4.1\nFactors Affecting Reserve Balances of Depository Institutions and\nCondition Statement of Federal Reserve Banks\nOctober 22, 2026\n1. Factors Affecting Reserve Balances of Depository Institutions\nMillions of dollars\n...",
    "markdown": "## FEDERAL RESERVE statistical release\n\nH.4.1\n\n## Factors Affecting Reserve Balances of Depository Institutions and\n\n...\n\n| Averages of daily figures |  |  |\n| --- | --- | --- |\n| Oct 21, 2026 | Oct 14, 2026 | Oct 22, 2025 |\n| 6,706,861 | - 789 | + 161,908 |\n...",
    "tables": [
        {
            "tableIndex": 0,
            "page": 2,
            "bbox": [251.75, 173.27, 503.75, 519.23],
            "rowCount": 33,
            "columnCount": 3,
            "header": ["Averages of daily figures", "", ""],
            "rows": [
                ["Averages of daily figures", "", ""],
                ["", "Change from week ended", ""],
                ["Week ended", "", ""],
                ["Oct 21, 2026", "Oct 14, 2026", "Oct 22, 2025"],
                ["6,706,861", "- 789", "+ 161,908"],
                ["6,470,860", "+ 492", "+ 187,540"]
            ],
            "markdown": "| Averages of daily figures |  |  |\n| --- | --- | --- |\n...",
            "strategy": "lines+text"
        }
    ],
    "chunks": [
        {
            "chunkIndex": 0,
            "page": 2,
            "charStart": 0,
            "charEnd": 608,
            "text": "FEDERAL RESERVE statistical release\nH.4.1\n...",
            "citation": { "type": "page", "page": 2, "label": "p. 2" }
        },
        {
            "chunkIndex": 1,
            "page": 2,
            "charStart": 460,
            "charEnd": 1659,
            "text": "...",
            "citation": { "type": "page", "page": 2, "label": "p. 2" }
        },
        {
            "chunkIndex": 2,
            "page": 2,
            "charStart": 1509,
            "charEnd": 2103,
            "text": "...",
            "citation": { "type": "page", "page": 2, "label": "p. 2" }
        }
    ],
    "ocrUsed": false,
    "charCount": 2103,
    "warnings": [],
    "extractedAt": "2026-10-23T05:39:57.921Z"
}

With includeCellBoxes: true each table also has "cells": [{"row": 4, "col": 0, "text": "6,706,861", "bbox": [251.75, 208.93, 337.25, 217.95]}, ...].

Coordinates are PDF points (1/72 inch) with the origin at the top-left corner of the page; width and height give the page size in the same unit, so dividing by them yields relative positions for any viewer.

In documents mode the item has pageCount, pagesProcessed, pages, metadata (title, author, creator, producer, creationDate), text (pages joined with \f), markdown, tables (all pages, each with page), chunks (all pages, each with page), ocrPagesUsed, charCount, warnings and extractedAt.

Family: the same chunks for PDFs and recordings

This Actor has a sibling, transcribe-to-json-chunks, which turns audio and video (MP3, WAV, M4A, MP4, MOV, WebM and more) into the exact same chunk schema with time citations instead of page citations. Both emit:

{ "chunkIndex": 0, "text": "...", "charStart": 0, "charEnd": 1187, "citation": { "type": "page", "page": 3, "label": "p. 3" } }
{ "chunkIndex": 0, "text": "...", "charStart": 0, "charEnd": 1187, "citation": { "type": "time", "startMs": 723000, "endMs": 768000, "label": "00:12:03-00:12:48" } }

So one ingestion pipeline can index chunks[].text with citation.label as the display reference, whether the source was a 40-page report or a 40-minute meeting recording. The chunker is the same code in both Actors (sentence-aligned, character-sized, with overlap), so chunk boundaries behave the same way too.

Pricing in plain words

Pay per page, nothing else: $0.003 per page processed (the page-processed event), plus $0.01 per page only when OCR actually ran on it (the ocr-page event), plus Apify's standard tiny Actor start fee. A 30-page report costs 9 cents; fully OCR'd it costs 39 cents. Both output modes charge per processed page. Failed downloads and unreadable PDFs are never charged. You set the maximum you are willing to spend per run; when the limit is reached the Actor stops parsing and still finishes successfully, and the SUMMARY record says so. Platform compute is included in the price.

Limits

• File size: 50 MB per PDF by default (maxFileSizeMb, up to 200). Up to 100 URLs per run.
• Pages: 50 per PDF by default (maxPagesPerPdf, up to 1000). pageCount always shows the true total.
• Text-layer PDFs only unless you turn OCR on. A scanned PDF returns empty text with ocr: "off"; use auto or always. OCR takes about 2 to 3 seconds per page and is English only (eng) in this build.
• Encrypted or password-protected PDFs are not supported and return an encrypted error item.
• Table detection is heuristic. Ruled tables are reliable. Tables drawn only with whitespace are found when they look like data (mostly numbers, dense); text-only tables without lines may be missed, and caption lines directly above an unruled table can be absorbed as a first row. Where a ruled table has sub-columns without lines, those values stay together in one cell. Check strategy and bbox when quality matters.
• Reading order follows the PDF's own layout; multi-column articles are read left to right by line, as most parsers do.
• URLs must be public. Pages behind a login (HTTP 401) or a bot wall (HTTP 403) return an error item; upload such files to storage you control (for example an Apify key-value store) and pass that URL.
• Runs are stateless. PDFs are fetched from the URLs you provide, parsed in memory and on the run's disk, and deleted when the run ends. Only the JSON you see in the dataset is kept.

Use it from an AI agent (Apify MCP server)

Connect your agent to https://mcp.apify.com (see the Apify MCP docs for Claude, Cursor, VS Code and others). Then:

1. Search for the tool: search-actors with the query "pdf to json" or "pdf table extraction".
2. Read the schema: fetch-actor-details for this Actor returns the input fields above and the dataset schema.
3. Call it: call-actor with {"pdfUrls": ["https://example.com/report.pdf"]}, or {"pdfUrls": ["..."], "pages": "1-10", "outputMode": "documents"} for one object per file, or {"pdfUrls": ["..."], "ocr": "auto"} when some pages may be scanned.
4. Read results with get-dataset-items. Index chunks[].text with page, charStart and charEnd as metadata; feed tables[].rows straight into a dataframe.

The Actor runs with limited permissions and charges only pay-per-event events, so it works with Apify's agentic payments.

Open-source credits and licenses

Everything inside is permissively licensed. No copyleft components are used or bundled.

Component	Role	License
pdfplumber	Text, words, characters and table detection	MIT
pdfminer.six	PDF parsing engine under pdfplumber	MIT
pypdfium2 (PDFium)	Page rendering for OCR	BSD-3-Clause or Apache-2.0
Pillow	Image handling for OCR	MIT-CMU (HPND)
pytesseract	Python bridge to Tesseract	Apache-2.0
Tesseract OCR with the eng model	OCR engine	Apache-2.0
Apify SDK for JavaScript, got-scraping	Actor runtime and downloads	Apache-2.0 / MIT
Base image apify/actor-node:22 (Node.js on Alpine Linux)	Runtime	MIT / various permissive

Python dependencies pulled in transitively (cffi, pycparser, cryptography, charset-normalizer, packaging) are MIT, BSD or Apache licensed. The Actor's own code is Apache-2.0.

FAQ

Does it work on scanned PDFs? Yes, set ocr to auto (mixed documents) or always (fully scanned). OCR pages cost $0.01 extra each. Only English OCR is installed today.

Why did a table come back with the row labels missing? When a PDF draws ruled boxes around the numbers but prints the labels outside the box, the label column is not part of the ruled table. The labels are still in text and markdown, and the table's bbox tells you where it sits on the page.

Can I get the position of every cell? Yes, set includeCellBoxes: true. Each table then carries cells: [{row, col, text, bbox}].

One item per page or one per document? outputMode: "pages" (default) gives one dataset item per page, which is the natural unit for RAG. outputMode: "documents" gives one item per PDF with every table and chunk tagged with its page.

Are my PDFs stored? No. Each file is downloaded from the URL you give, parsed, and deleted when the run ends. Results live in your own dataset.

What does it cost? $0.003 per page processed, plus $0.01 per page only when OCR runs. A 30-page report with a text layer costs 9 cents.

Where do recordings fit? The sibling Actor transcribe-to-json-chunks returns the same chunk schema with time citations. Working pipelines for both are in the rag-ingestion-examples repository.

Support

Open an issue on the Actor's Issues tab in Apify Console. Include the run id, the PDF URL (if you can share it) and the input you used.