Convert PDF Documents to Markdown

Why scanned PDFs produce nothing (and what to use instead)

A scanned document is a PDF whose pages contain a single raster image. There is no text layer at all — no show-text operators, no glyphs, no font information. Running text extraction against it returns an empty string, and no amount of parser sophistication changes that, because the information simply is not present in the file.

Recovering text from an image requires optical character recognition: a model that looks at pixels and predicts characters. That is a fundamentally different and much heavier operation, and it is not something this converter performs. If your PDF opens in a viewer but you cannot select or search the text with Ctrl+F, it is a scan, and you need an OCR tool rather than a text extractor.

A useful diagnostic: open the PDF and try to select a single word. If the selection highlights a rectangle covering the whole page rather than the word, the page is an image. Many PDFs are also hybrids — a born-digital cover page followed by scanned appendices — in which case extraction returns the first pages and silently skips the rest, which is why the converter warns when output is suspiciously short relative to page count.

Why tables do not survive PDF extraction

There is no table object in a PDF's content stream. What looks like a table is a set of text runs positioned to line up, sometimes accompanied by unrelated line-drawing operators for the borders. Nothing marks a cell as a cell, associates a value with a column header, or records that two visually adjacent numbers belong to the same row.

Reconstructing a table therefore means inferring a grid from x-coordinate clustering, and that inference is fragile in exactly the cases tables are used: right-aligned numeric columns, cells whose content wraps onto two lines, merged header cells spanning several columns, and columns separated by whitespace rather than rules. A converter that emits confident-looking pipe tables from PDF geometry will produce plausible but subtly wrong data — cells shifted by one column, wrapped values split across rows — which is considerably worse than producing plain text, because the corruption is not visible.

This converter therefore does not attempt table reconstruction. Tabular regions come out as positioned text lines. If your source data originated in a spreadsheet, converting the original file with the Excel or CSV converter will give you a correct Markdown table; going via PDF is a lossy round-trip that discards the structure you need.

Streaming extraction and constant memory use

Large PDFs are processed one page at a time in a sequential loop rather than in parallel. This is a deliberate architectural choice with a measurable consequence: after each page's text is extracted, the page object's cleanup routine runs and the reference is dropped, so the browser can reclaim that memory before the next page loads.

The obvious alternative — mapping every page to a promise and awaiting them all at once — is faster on small documents and fails badly on large ones, because it holds every decoded page resident simultaneously. Peak memory then scales linearly with page count, and a few hundred pages is enough to trigger an out-of-memory crash in a browser tab. Under sequential extraction, peak memory stays at roughly one page regardless of whether the document is five pages or five hundred.

Because extraction runs inside a Web Worker, the main thread stays responsive throughout: the page scrolls, buttons respond, and a determinate progress bar reports "page N of TOTAL" as the loop advances. A long PDF takes longer, but it never freezes the interface and never hits a size ceiling imposed by an upload limit.

Encryption, permissions and password-protected files

PDF supports two distinct kinds of protection, and they behave differently. An owner password restricts permissions — printing, copying, editing — while leaving the content readable; a user password encrypts the document so it cannot be opened at all without the passphrase.

Encrypted PDFs cannot be parsed without the password, so they must be unlocked in a PDF viewer and re-saved before conversion. Permission-restricted files can usually be read, but bear in mind that a copy-protection flag is a statement of intent by the document's author: extracting text from a document you are not licensed to reuse is a legal question, not a technical one, and the converter has no way to make that judgement for you.

Choosing the right method for your PDF

Browser-based extraction is the right tool when the PDF is born-digital, the text layer is intact, and privacy matters — which covers the large majority of contracts, reports, papers and specifications. It is the wrong tool for scans, for documents whose value is in their tables, and for batch pipelines over thousands of files.

The matrix below compares the four realistic approaches on the dimensions that actually drive the decision. For a longer walkthrough of when each one is appropriate, including the cost and accuracy trade-offs of OCR and vision models, see the dedicated guide linked further down this page.