How to Convert PDF to Markdown for LLMs & RAG

PDFTEQ Team Engineers at PDFTeq

Published Oct 18, 2026 Updated Oct 18, 2026

8 min read

Share

• Twitter / X
• LinkedIn

• ---

• Copy Link

If you have ever pasted a 20-page research paper or financial quarterly report into an AI chat window or ingestion pipeline, you know the frustration. The model hallucinates key figures because numbers from Table Column 1 were slammed into Column 3, or it misattributes quotes because a two-column layout was read left-to-right across the horizontal gutter.

Markdown has become the lingua franca of large language models. Frontier models like GPT-4o, Claude 3.5 Sonnet, and Gemini 1.5 Pro are natively trained on trillions of tokens formatted in Markdown. When text arrives with explicit header markers (#), bullet points (-), and GitHub Flavored Markdown (GFM) pipe tables (| Col | Col |), reasoning accuracy jumps and chunking becomes deterministic.

1. Why LLMs and RAG Pipelines Demand Markdown

Large language models process text as a sequential stream of tokens. While modern frontier models boast context windows from 128k to 2 million tokens, context quality matters far more than context volume. When raw, unformatted text enters the context window, several critical failure modes emerge:

2. The Anatomy of a PDF: Why Extraction Fails

To understand why simple copy-pasting fails, you must understand how Adobe designed the PostScript and Portable Document Format (PDF) specification. A PDF is fundamentally a vector drawing instruction set. When a page renders, the internal content stream looks like this:

Notice what is missing: there is no <paragraph> tag, no <table> element, and no column boundary. The PDF engine simply paints character strings at absolute 2D Cartesian coordinates (x, y). A human eye groups these elements visually using Gestalt proximity principles, but a machine reading the stream in file order or strictly top-to-bottom can easily produce gibberish.

3. Solving the Multi-Column Reading Order Problem

Academic journals (such as IEEE, ACM, Nature, and arXiv preprints), legal filings, and newspapers frequently format text into two or three vertical columns. When simple extractors read a page with two columns, they sort text elements solely by vertical position. Because Line 1 of Column 1 and Line 1 of Column 2 share roughly the same vertical baseline, the text is merged horizontally across the column gutter.

PDFTeq’s client-side converter inspects the distribution of text bounding boxes across the page width. If a distinct horizontal empty band (a gutter between 38% and 62% of page width) is detected with substantial text on both sides, the algorithm partitions the page: any full-width title or abstract at the top is parsed first, followed by the left column read top-to-bottom, followed by the right column read top-to-bottom, and finally any full-width footnotes or bottom spanning blocks.

4. Hyphenation Repair and Tabular Alignment

Two subtle issues severely degrade embedding quality in vector databases:

End-of-Line Hyphenation Repair

Typesetting engines routinely insert hyphens to justify right margins. When a word like multi-column or con- verter is broken across lines, standard copy-pasting yields either broken tokens or missing words. Our converter inspects adjacent lines within the same vertical block: if a line ends with an alphabetical character followed by a hyphen (e.g. [A-Za-z]-) and the next line begins with a lowercase letter, the hyphen is removed and the word is seamlessly re-joined.

Tabular Alignment via Heuristic Clustering

Tables in PDFs rarely contain borderlines that programmatic text extractors can read. Instead, they are whitespace-aligned. Our converter detects consecutive lines that share multiple common horizontal tab stops and format them as Markdown pipe tables with a clean separator row (|---|---|). If a table has clear numeric columns, they remain structured and queryable by LLMs.

5. Tool Comparison: Browser vs Python (Docling, Marker, MarkItDown)

Different document ingestion workflows require different tools. Here is an honest, objective breakdown of how PDFTeq compares to popular command-line and machine-learning frameworks:

When to use PDFTeq: You need an immediate, zero-friction conversion of financial statements, agreements, business reports, or articles without sending sensitive data over the wire or spinning up a Python virtual environment.

When to use Docling or Marker: You are ingesting thousands of scanned academic papers containing complex multiline LaTeX equations, embedded chemical structures, or heavily nested matrices.

6. Step-by-Step: Converting PDFs Locally on PDFTeq

Converting your PDF to Markdown takes less than five seconds using PDFTeq:

7. Best Practices for RAG Chunking and Token Management

Once you have converted your PDF to Markdown, how should you store and index it in your vector database (such as Pinecone, Qdrant, Chroma, or pgvector)? Here are the recommended industry patterns:

1. Header-Aware Semantic Chunking

Instead of naive fixed-window token slicing (e.g. chunking every 500 tokens with 50-token overlap), use LangChain's or LlamaIndex's MarkdownHeaderTextSplitter. Configure it to split on #, ##, and ### tags. This guarantees that each chunk corresponds to an entire logical concept, keeping methods, findings, and prerequisites together.

2. Preserve Metadata in Chunk Headers

Prepend the document title and page number to the top of each vector chunk. For example:

This allows your retrieval engine to cite specific page numbers back to the user when generating answers.

8. Frequently Asked Questions

Written by the PDFTEQ Team

We build and maintain PDFTeq's suite of secure, browser-first PDF utilities. Our tools run strictly on the client side using WebAssembly and modern JavaScript, ensuring private documents never leave your device.

More technical guides