PDF to markdown for LLMs, and every other file agents are handed
Most enterprise questions are answered by documents, not web pages. CrawlGenie parses them without a model: PDF text layers and OCR for scans, Office and OpenDocument files, images, Google Docs, Sheets and Slides links, YouTube caption tracks. Ask a question and get the passages; give a page range and skip the rest.
Which document types are supported?
| Source | How it is read | Notes |
|---|---|---|
| PDF with text | pdf.js text extraction, lines by position, tables rebuilt as pipe rows | pages: "1-5,9" to limit work |
| Scanned PDF, images | OCR with Tesseract on the server (auto when the text layer is empty) | Confidence reported; runs inside the platform, never sent to a third-party model |
| DOCX, XLSX, PPTX, ODT, ODS, ODP, RTF | Native zip/XML parsing, no Office needed | Sheets become rows; slides become sections |
| Google Docs, Sheets, Slides, Drive, Dropbox, GitHub links | Rewritten to export or raw URLs | Public or "anyone with the link" |
| YouTube | Caption track → timestamped paragraphs | No speech model; videos without captions return "no transcript" |
How do I turn a PDF into markdown for an LLM?
For developers: the exact call
POST /v1/parse {"url": "https://example.com/annual-report.pdf"} # whole document
POST /v1/parse {"url": "...pdf", "q": "revenue by segment", "pages": "30-60"} # passages from pages 30–60
POST /v1/parse {"base64": "", "filename": "scan.png", "ocr": "auto", "redactPII": true}
POST /v1/scrape {"url": "https://docs.google.com/spreadsheets/d//edit#gid=0", "q": "total for March"} How is this cheaper than uploading the file to ChatGPT or Claude?
A file attached in the chat is processed and billed by the platform, in full, every time. Through the connector, CrawlGenie parses it once and sends the model only the passages that answer, typically 5–15% of the document, and you can save it to a knowledge collection so the next question costs a few hundred tokens.
What is OCR and when does it run?
OCR (optical character recognition) turns pixels into characters. A scanned invoice has no words a program can read until OCR runs. CrawlGenie renders each page, runs OCR inside the platform, keeps a confidence score, and triggers automatically when a PDF's text layer is empty or sparse. It is slower than reading a text layer (about 1–3 seconds per page) and never sent to a third-party model.
What happens to personal data?
redactPII replaces emails, phone numbers, Luhn-valid card numbers, IBANs, US SSNs and public IPs with placeholders before the text leaves the server, and reports counts. Names and street addresses are not detected; the docs say so rather than imply otherwise.
Related: documents guide · knowledge base · web scraping for LLMs · use cases by industry.
What would this save you?
Move the sliders. The saving is computed from the measured ratios above and the model price you pick.
Assumes 10% of the page is kept per read (measured range 5–15%). Receipts in the product show the real number for your traffic.
Questions people ask
Can it read Word and Excel files without Microsoft Office?
Yes. DOCX, XLSX and PPTX (and ODT, ODS, ODP) are zip files with XML inside; CrawlGenie parses them directly.
Does it transcribe audio or video?
No. YouTube works through caption tracks. Audio and video files are listed by the page_media extractor so you can route them to a speech service you choose.
How large a PDF can it handle?
Hundreds of pages; use pages to read only a range, and q to get passages instead of the whole text. Documents are capped at 25 MB per fetch by default.
Spend tokens on answers, not on pages.
Create a workspace in a minute. Connect ChatGPT, Claude, Cursor or your own agents. Every answer comes with a token receipt.