MarkdownPDF

Split a PDF for NotebookLM, ChatGPT and Claude

A textbook, a thesis or a year of board minutes rarely fits in one upload. This tool counts the words, reads the chapters, and cuts the PDF into parts each tool will actually accept — in your browser, with nothing uploaded.

Drop the PDF to split, or browse

Processed on your device — never uploaded

The upload limits, and where they come from

Checked against each vendor's own documentation on 2026-09-18. Limits change without notice — every row links to the page it came from, so you can confirm it yourself.

ToolPer-file limitWorth knowing
NotebookLM500,000 words or 200 MB per sourceUp to 50 sources per notebook on the free plan; paid plans allow more. NotebookLM Help: Add or discover new sources
ChatGPT512 MB and 2 million tokens per file(token cap converted to words — an estimate)The token cap is converted to about 1.5 million words, an estimate for English text. OpenAI Help Center: File Uploads FAQ
Claude (chat)Charts and images are read only in PDFs of 100 pages or fewerClaude accepts PDFs up to 1,000 pages and 500 MB in a chat, but beyond 100 pages it reads the text only. Parts of 100 pages keep every chart and figure visible to it. Claude Help Center: Upload files to Claude
Claude Projects30 MB per file in a Project Claude Help Center: Upload files to Claude
Gemini100 MB per file, up to 10 files per promptGemini's default context holds roughly 1,500 pages of text in total. Gemini Apps Help: Upload & analyze files

How it works

  1. 1

    Drop the PDF

    The tool reads it locally: page count, words per page, file size, and the chapters in its bookmarks. Nothing is sent anywhere.

  2. 2

    Pick the AI tool

    Each one has its own limits. The plan updates as you choose, and tells you straight away when no split is needed at all.

  3. 3

    Create the parts

    Each part is a real PDF you can download one by one or as a ZIP. A part that still comes out too large is split again automatically.

Split by chapter, not by page count

Cutting a book every 80 pages puts the end of chapter 3 and the start of chapter 4 in the same file, and splits an argument across two sources. A model answering from part 2 has the conclusion without the premise. When a PDF carries bookmarks — most textbooks, reports and theses do — this tool uses them, so each part is a self-contained section with a name you recognise in the answer's citation.

It also tells you when to do nothing. Plenty of documents already fit: a 300-page report is usually well under NotebookLM's 500,000 words. Splitting it anyway would only spend source slots you may need for other documents.

Markdown often beats splitting

Limits are counted in words, tokens or megabytes — and a PDF is a heavy way to carry text. The same document as Markdown is a fraction of the size, keeps its headings, and drops the layout noise that confuses extraction. If your goal is answers rather than page fidelity, convert first with the PDF to Markdown converter (with OCR for scans) and split only if it is still too large. The comparison of what models actually receive shows why.

Frequently asked questions

Why split a PDF before uploading it to an AI tool?

Because every tool has a per-file ceiling — words, tokens, megabytes or pages. Past it, the upload is refused or the file is silently truncated, and you get confident answers based on half a document. Splitting keeps every page inside the window, and makes it obvious which part an answer came from.

Is my PDF uploaded anywhere?

No. The analysis (pdf.js) and the splitting (pdf-lib) both run in your browser. The file never leaves your device, which matters for the contracts, medical files and unpublished research people usually need to split.

What does “split at chapter boundaries” do?

If the PDF has bookmarks, each top-level bookmark starts a new part, so a part is a chapter rather than an arbitrary page range. Chapters that are still over the limit are subdivided, and anything before the first chapter becomes a “front matter” part. Without bookmarks, the tool packs as many consecutive pages as each limit allows.

Why does Claude split at 100 pages when it accepts 1,000?

Because of what it reads, not what it accepts. Claude analyses charts, diagrams and images in PDFs of 100 pages or fewer; from 101 to 1,000 pages it processes the text only. Parts of 100 pages keep the figures visible to it. Use custom limits if you only care about text.

My PDF is scanned — will this help?

Splitting does not create text. A scanned page stays an image, and the AI tool has to run its own OCR, with results you cannot inspect. Converting to Markdown with OCR first gives you text you can check, and usually better answers.

Are these limits up to date?

They were checked against the vendors' own documentation on 2026-09-18, and each row links to its source. These limits change: if you hit one that does not match, use the custom limits and tell us so the page can be corrected.

Limitations

Related tools

Guides for AI document workflows