PDF to Markdown

Plain text throws away the one thing an LLM finds most useful about a document: its shape. Markdown keeps that shape in a lightweight, model-friendly way.

A heading written as ## Section 3 tells a model — and a retrieval pipeline — exactly where one topic ends and the next begins. A bulleted list rendered as real - items reads as a set of parallel points rather than an ambiguous run-on. That structure is what lets you chunk a document on its headings instead of blindly every N characters, so each chunk you embed is a coherent unit of meaning rather than a fragment sliced mid-sentence.

Markdown is also close to the format most models were heavily trained on, so headings, lists, and emphasis are interpreted the way you would expect. And unlike HTML or a raw PDF, Markdown spends almost no tokens on markup — you get the structure without paying for a pile of tags in your context window. If you only need selectable prose with no structure, our PDF to Text tool is the lighter choice; choose Markdown when the hierarchy matters.

A PDF does not actually store headings and lists. It stores glyphs at coordinates on a page. Rebuilding structure means inferring intent from those clues, and that inference is best-effort, not exact.

Here is what the tool does well: it reads font sizes and weights to promote likely titles and section headers to Markdown headings, detects bullet and number prefixes to form lists, preserves paragraph boundaries, and keeps reading order correct on single-column pages almost every time.

Here is where it approximates. Multi-column layouts — papers, newspapers, brochures — have to be linearized into one stream, and the tool follows the order recorded in the PDF, which usually but not always matches how you would read it. Complex tables are the hardest case: simple grids convert to Markdown tables, but merged cells, nested headers, and heavily designed tables are flattened to their best readable approximation. Heading levels are inferred from typography, so a document with unusual styling may get the nesting slightly wrong. For LLM context this is normally fine — models are tolerant of minor structural noise — but it is worth a glance before you rely on the output verbatim.

This tool exists because "turn a PDF into something I can feed a model" has become an everyday task, and most PDFs are hostile to it out of the box.

For ad-hoc context, the workflow is direct: convert, copy the Markdown, and paste it into your chat. You get clean structure and no wasted tokens, and because the file never left your machine, you can do it with documents you would never upload to a third-party service.

For RAG, the Markdown gives you natural split points. Chunk on #/## boundaries so each embedded passage is a self-contained section; keep the heading text in the chunk so retrieval has a topical anchor; and store the .md alongside your source PDF for traceability. Cleaner input here directly improves retrieval quality — garbage-extracted PDFs are a common, invisible cause of bad RAG answers.

When you outgrow converting files by hand, a metered REST API and a hosted MCP server are on the roadmap, so the same file-to-Markdown conversion becomes a call you make from a script, an ingestion job, or an agent. The browser tool and the API are meant to be two front doors to the same conversion.

Most online converters send your PDF to their servers to do the work, which means a copy of your document lands in someone else's storage, however briefly. This tool does the parsing in your browser instead, using the same local runtime that powers offline web apps.

The consequences are concrete. There is no upload, so there is no server-side copy of your document, no processing log tied to your file, and nothing for a future breach or subpoena to reach. It also means the tool keeps working on the documents that matter most — the confidential ones you are contractually or ethically barred from pasting into a cloud service. The first conversion loads a small amount of parsing code, then everything runs locally and works offline afterward. If you want proof rather than a promise, watch the network tab: the conversion finishes with no file leaving your browser.