Convert PDF to Markdown — Clean Output for LLMs & RAG

Why Convert PDFs to Markdown?

If you've ever pasted a PDF into ChatGPT, Claude, or your local Llama setup and gotten garbage back, you've already learned the lesson the hard way: PDFs are a terrible input format for language models.

PDFs store text as positioned glyphs — `(x: 142, y: 718) "T"`, `(x: 148, y: 718) "h"`, `(x: 152, y: 718) "e"` — not as flowing paragraphs. Naive PDF-to-text extraction gives you:

Columns concatenated wrong

Headers and footers interleaved with body text

Footnotes embedded in mid-sentence

Lists turned into walls of text

Tables flattened into one cell per line

Math equations replaced with question marks

For RAG (retrieval-augmented generation), this destroys retrieval quality because chunk boundaries fall in the wrong places and embeddings represent positional noise instead of semantic structure.

Markdown solves all of this.

What Good PDF→Markdown Looks Like

A typical research paper input/output:

Input PDF (raw text extraction): ``` Abstract Introduction Recent advances in transformer 1 architectures have shown that self attention is a Methods more efficient mechanism than recurrence... ```

Output Markdown: ```markdown

Abstract

Recent advances in transformer architectures have shown that self-attention is a more efficient mechanism than recurrence...

1. Introduction

[paragraph here]

2. Methods

[paragraph here] ```

The Markdown version is chunkable, embeddable, and directly usable in any LLM pipeline.

Use Cases

1. Feeding Internal Docs to ChatGPT/Claude

Paste Markdown — not PDF — into the chat. Token cost drops 30–60% (PDF text extraction includes a lot of garbage). Quality of answers goes up dramatically.

2. RAG Pipeline Ingestion

LangChain, LlamaIndex, Haystack, and Chroma all ingest Markdown natively:

```python from langchain.text_splitter import MarkdownHeaderTextSplitter

splitter = MarkdownHeaderTextSplitter(headers_to_split_on=[ ("##", "section"), ("###", "subsection"), ]) chunks = splitter.split_text(markdown_content) ```

Chunks split on heading boundaries — exactly what you want for retrieval.

3. Obsidian / Notion / Logseq Notes

Drop a PDF you want to take notes on, convert to Markdown, paste into your note-taking app. All headings become an outline, all links become wiki-links.

4. Static Site Publishing

Hugo, Jekyll, MkDocs, Astro, Next.js — they all consume Markdown. Your PDF technical docs become a searchable web site in 10 minutes.

5. Translation / LLM Editing

Round-trip a PDF through Markdown → LLM rewrite → Markdown → back to PDF (using Markdown to PDF) for a clean, edited output.

Why "No Upload" Matters Here Especially

PDFs you're feeding to LLMs are often the most sensitive things you own:

Internal product specs that competitors would love

Research drafts before publication

Customer data in compliance reports

Financial documents in audits

Legal exhibits in active matters

Uploading them to a SaaS conversion service to "make them LLM-friendly" — and then uploading the output to ChatGPT/Claude/Gemini — doubles your exposure surface. ExactPDF's PDF to Markdown runs entirely in your browser, so the conversion step adds zero risk.

What's Preserved (and What Isn't)

Element	Preserved	Notes
Headings	✅	Inferred from font-size hierarchy
Paragraphs	✅	Reflowed correctly across columns
Bullet lists	✅	Detected from bullet glyphs (•, –, ●)
Numbered lists	✅	Detected from "1.", "1)", "(1)" patterns
Links	✅	Embedded URLs preserved as Markdown links
Tables	✅	Output as GitHub-flavored Markdown tables
Code blocks	✅	Detected by monospace font + indentation
Inline code	⚠️	Sometimes misclassified (font heuristic)
Bold / italic	✅	Detected from font weight/style
Math equations	⚠️	Output as raw text — won't render unless you wrap in \$\$
Images	⚠️	Extracted as separate files (Markdown references them)
Footnotes	✅	Output as superscript + bottom-of-section reference
Headers / footers	✅ removed	Page numbers and running titles dropped
Multi-column layout	✅	Re-flowed into single-column reading order

Step-by-Step: PDF to Markdown for RAG

Step 1: Convert

Open PDF to Markdown, drop your PDF in, click Convert. Output appears in seconds — no upload.

Step 2: Inspect

Scroll the output. Check:

Headings hierarchy looks right

Tables are intact

Code blocks are fenced (`\`\`\``)

Step 3: Save and chunk

Download the .md, then in your RAG pipeline:

```python from langchain.document_loaders import UnstructuredMarkdownLoader from langchain.text_splitter import MarkdownHeaderTextSplitter

loader = UnstructuredMarkdownLoader("paper.md") docs = loader.load() splitter = MarkdownHeaderTextSplitter(...) chunks = splitter.split_text(docs[0].page_content) ```

Step 4: Embed and index

Pass chunks to your embedding model (OpenAI text-embedding-3, Cohere embed-v3, BGE, Jina, whatever) and into your vector store.

Step 5: Retrieve

At query time, retrieve top-k chunks, pass to your LLM as context. Quality is dramatically higher than chunking raw PDF text.

Scanned PDFs

If your source is a scanned image PDF (no text layer), the Markdown converter has nothing to work with. Two-step pipeline:

OCR PDF → adds text layer using Tesseract.js (in browser)

PDF to Markdown → converts the text-layered PDF to Markdown

Both run client-side, so your scanned document never leaves your laptop — important for medical records, legal exhibits, and historical archives.

Comparison

Tool	Server upload?	Quality	Cost
ExactPDF PDF to Markdown	No	High (heading/table aware)	Free
Online "PDF to Markdown" sites	Yes	Variable	Often paid
Pandoc local	No	Medium (lossy on tables)	Free, install needed
LlamaParse / Unstructured.io	Yes (API)	High	Pay per page
Adobe API	Yes (API)	High	Enterprise pricing
Manual conversion	No	Highest	Slow + expensive

FAQ

Why convert PDFs to Markdown?

Markdown is the cleanest input format for LLMs, RAG pipelines, note-taking apps, and static-site generators. PDF's positioned-glyph format adds noise that hurts retrieval quality.

Does it preserve tables and code blocks?

Yes — tables become GFM tables, code blocks are detected by font and indentation, headings are inferred from font-size hierarchy.

Is it good enough for RAG ingestion?

Yes. The output is structured Markdown ready to chunk and embed. Chunk on heading boundaries (split on ##/###), keep tables as units.

Does the PDF get uploaded?

No. Conversion happens locally in your browser.

Can I convert scanned PDFs to Markdown?

Run OCR PDF first to add a text layer, then run PDF to Markdown.

Try PDF to Markdown — Free, No Upload →