Turn web pages into clean Markdown for AI and RAG: one row per URL with the main content, title, author, publish date and the page's links split into same-site and external; also works on pages that need JavaScript; pages that cannot be read say why and are not billed. Not affiliated with the websites read
PDF to Markdown
Turn PDF links into clean Markdown for AI and RAG: one row per PDF with headings, paragraphs, lists and tables kept as Markdown tables, multi-column pages read in order, running headers and page numbers removed, plus title, author, page count and creation date. Billed per PDF converted, no start fee; links that are not PDFs, scanned PDFs without text, and blocked downloads get a status row and are not billed. Not affiliated with the sites the files come from
Aperçu
Give it a list of PDF links and get back one row per PDF: the whole document as clean Markdown, plus its title, author, creation date and page count.
Reports, papers, filings, manuals and datasheets still live in PDFs, and most of them are hard for an AI to use as they are: text comes out in the wrong order on multi-column pages, tables turn into a stream of numbers, and every page repeats the same header and footer. This app rebuilds the document as Markdown — headings, paragraphs, lists and tables — so a model can summarize it, answer questions from it or store it in a knowledge base.
Every link gets a row with a clear status: converted, not a PDF, a scan without text, password protected, refused by the site, gone, or not reached because the run was cancelled or ran out of time. You never send 10 links and get 7 rows back without knowing what happened to the other 3, and links that were not converted are not charged.
Data notes
- This app downloads the files at the links you give and converts their format; the documents belong to their publishers, so follow their copyright and terms of use. This is not an official product of any of these sites and is not affiliated with them. The fee covers downloading and conversion, not the documents themselves.
- Tables drawn with ruling lines become Markdown tables. Tables without lines (common in financial statements) come out one line per row, with the cells separated by " | ", so each label stays next to its numbers.
- Pages with two or three columns are read column by column. Running headers, footers and page numbers are removed.
- Headings are recognised from larger or bold section titles; "Title", "Author" and "Created" come from the file's own properties and are left empty when the file does not set them.
- Long PDFs can be read in parts: a row says which page to continue from when the file has more pages than were read.
- Not covered: scanned PDFs without a text layer (no text recognition is done), password-protected files, files over 50 MB, and files behind a login. Images and charts are not described; only their text labels may appear.
What the results look like
Each row is one link.
| Link | Status | Title | Pages read | Tables | Characters |
|---|---|---|---|---|---|
| https://arxiv.org/pdf/1706.03762 | ok | Attention Is All You Need | 15 | 0 | 38743 |
| https://nvlpubs.nist.gov/nistpubs/CSWP/NIST.CSWP.29.pdf | ok | The NIST Cybersecurity Framework (CSF) 2.0 | 32 | 2 | 66146 |
| https://example.com | not_pdf | 0 | 0 | 0 |
Each row also has the Markdown content, the author, creation date, total pages in the file, start page, file size, status details and the time it was read.
Use cases
- Feed an AI: hand "Content (Markdown)" to a model to summarize a report, compare two filings or answer questions about a manual.
- Build a knowledge base: store "Content (Markdown)" with "Title" and "Link" for each document, split it on its headings, and cite the source in answers.
- Pull figures out of reports: tables arrive as rows and columns, so a model or a script can pick out the numbers it needs.
- Clean a document list: filter on "Status" to find links that are dead, refused, scanned or not PDFs at all, and fix or drop them.
Périmètre et limites
Up to 100 PDF links per run and up to 500 pages per PDF (100 by default, from start_page); files over 50 MB are not downloaded. Only PDFs with a text layer are converted: scanned PDFs come back as no_text_layer, no text recognition (OCR) is done. Files are downloaded directly, robots.txt is respected, logins and paywalls are never bypassed. Markdown is cut at about 900 KB per row. Billed per PDF converted; rows for links that were not converted are free. The fee covers downloading and conversion, not the documents themselves
- Adapté pour
- You have links to PDF files (papers, annual reports, filings, government reports, manuals, datasheets) and need their text as Markdown for an LLM to read, summarize or answer from, or to load into a RAG knowledge base
- You need tables from PDFs kept as rows and columns rather than jumbled text
- A PDF is longer than your model's context: read it in parts with max_pages and start_page
- À ne pas utiliser pour
- The link is a web page, not a PDF file: use a web page to Markdown app (such links come back as not_pdf)
- The PDF is a scan or photo without a text layer: no text recognition is done here (it comes back as no_text_layer, not billed)
- The file sits behind a login or a paywall, or needs a password
Gestion des échecs
Comportement en cas d'échec et de nouvelle tentative déclaré par le développeur. Nous vous recommandons de l'inclure dans votre prompt système lors de l'intégration.
- 1Every link gets exactly one row, in input order. `status` says what happened; only ok is billed (`billed_pdfs` = 1)
- 2blocked, timeout and error can be retried later; robots_disallowed means the site's robots.txt asks automated readers not to fetch the file
- 3When `pages_read` is less than the pages left in the file, `status_text` says which start_page to use next; `truncated` = true means the Markdown hit the row size limit, so read fewer pages per call
Entrée
Paramètres requis pour appeler cette App, générés à partir de l'input.schema du manifest.json.
- PDF links
urlsRequisarray<string>Links to PDF files, 1–100. Each link gets one row, in the order given. A link given twice is read once (the second row says duplicate).Nombrejusqu'à 100 élémentsExemple["https://arxiv.org/pdf/1706.03762"] - Max pages per PDF
max_pagesFacultatifintegerHow many pages of each PDF to read, counted from start_page; 1–500. Pages after that are not read, and the status details say where to continue.Plage1–500Défaut100 - Start page
start_pageFacultatifintegerThe first page to read (1 = the first page). Use it to continue a long PDF in a second call; it applies to every link in the run.Plage1–Défaut1
Sortie
Structure des champs d'un enregistrement, générée à partir de l'output.schema du manifest.json.
- Link
input_urlstringThe link as given in the input, to match rows to inputsExemplehttps://arxiv.org/pdf/1706.03762 - Final link
final_urlstringThe address the file was downloaded from; differs from the input when the link redirectedExemplehttps://arxiv.org/pdf/1706.03762 - Status
statusstringok = converted; not_pdf = not a PDF (often a web page); no_text_layer = scan without text; encrypted = password protected; unreadable = damaged file; too_large = over 50 MB; page_out_of_range = start_page is past the last page; robots_disallowed = robots.txt disallows it; not_found = HTTP 404/410; blocked = site refused the download; timeout; dns_error; error; invalid_url; duplicate = same link earlier in the run; not_reached = the run was cancelled or timed out first. Only ok is billedExempleok - Status details
status_textstringPlain-English explanation of the status, with the reason when nothing was converted; also says which pages were read when the PDF has more pages, and when the Markdown was cut at the record size limitExempleConverted
Exemple de données
Généré à partir d'output.schema. Les champs sans valeur d'exemple utilisent des espaces réservés structurels.
{
"input_url": "https://arxiv.org/pdf/1706.03762",
"final_url": "https://arxiv.org/pdf/1706.03762",
"status": "ok",
"status_text": "Converted",
"billed_pdfs": 1,
"title": "Attention Is All You Need",
"author": "<author>",
"created_at": "2024-04-10T21:11:43Z",
"page_count": 15,
"start_page": 1,
"pages_read": 15,
"tables_found": 0,
"markdown": "Provided proper attribution is provided, Google hereby grants permission to reproduce the tables and figures in this paper solely for use in journalistic or scholarly works.\n\n# Attention Is All You Need\n\n...",
"char_count": 38757,
"truncated": false,
"file_size_bytes": 2215244,
"checked_at": "2026-11-17T03:53:42+00:00"
}Intégration
Cette App peut être intégrée via MCP, API, SDK ou export de fichiers — tous les canaux partagent les mêmes capacités et la même tarification. Chaque requête s'authentifie avec l'en-tête Authorization: Bearer à l'aide d'une API Key (longue durée, créée dans la console Data Hub) ; les clients MCP peuvent aussi se connecter via OAuth, sans clé. D'autres options comme la CLI et Skill arrivent bientôt. Cette App s'exécute de manière asynchrone et peut être annulée en cours de route : Après l'annulation, l'exécution s'arrête. Les enregistrements déjà livrés sont conservés et facturés.
Avec MCP (Model Context Protocol), vous pouvez appeler cette App directement depuis des clients IA comme Claude et Cursor. Choisissez votre client et votre mode d'authentification, puis copiez la config ci-dessous.
Config du client
Remplacez la valeur après Bearer par votre API Key longue durée. Fonctionne dans n'importe quel client, en CI ou en environnement sans navigateur.
{
"mcpServers": {
"octoparse_datahub": {
"type": "http",
"url": "https://mcp-v2.octoparse.com?pin=MagicData/pdf-to-markdown",
"headers": { "Authorization": "Bearer <YOUR_API_KEY>" }
}
}
}Laissez l'IA tout configurer pour vous
Vous ne voulez pas modifier les configs à la main ? Copiez le prompt d'installation et collez-le dans n'importe quel client IA — il effectuera la configuration à sa manière. (Par défaut, il vous guide pour vous connecter et ne vous demande jamais d'envoyer votre API Key dans la conversation.)
detail.access.mcp.composeHint
Tarifs
Mesuré en additionnant le champ de sortie billed_pdfs. Les exécutions échouées ne sont pas facturées.
Plusieurs événements de facturation se cumulent indépendamment. Voir chaque élément pour les détails. Les exécutions échouées ne sont pas facturées.
Voir les créditsEssayer maintenant
Renseignez les paramètres et lancez l'exécution — les résultats proviennent d'un appel réel.