How to Convert PDF to Markdown for ChatGPT, Claude & RAG (2026)

If you've ever pasted a raw PDF into ChatGPT and gotten back broken answers, mangled tables, or missing formulas — it's because LLMs don't understand PDF. They understand structured text, and Markdown is the lightest structured format every LLM (ChatGPT, Claude, Gemini, Llama) reads deeply: headings, lists, tables, code, LaTeX formulas.

Converting PDF to Markdown is a non-negotiable preprocessing step for any serious AI workflow:

• Document Q&A with ChatGPT/Claude — bots answer more accurately when the input is Markdown, not flat pasted text.
• RAG (Retrieval-Augmented Generation) — embedding/chunking pipelines need block lists with metadata (bbox, block type, page) for accurate retrieval.
• OCR scanned PDFs into Markdown — contracts and forms only become useful when text preserves diacritics and clear heading structure.

The BetaPDF PDF-to-Markdown tool runs Premium vision AI on dedicated GPU, processing 9 pages in just 22-30 seconds and returning both .md + .json bbox in one ZIP.

Step 1: Upload your PDF (max 50 pages)

Open BetaPDF PDF to Markdown, click Choose PDF file, or drag-and-drop into the upload zone. The tool accepts both digital PDFs (exported from Word, LaTeX) and scanned PDFs (photographs, scanned contracts).

Note: if your file exceeds 50 pages, use Split PDF to break it up first.

Step 2: Pick language + formula/table options

Choose a Language — "Auto (multilingual)" works for 95% of cases, including mixed Vietnamese-English documents. Toggle Extract math formulas on for documents with formulas (academic papers, exam sheets) — output is LaTeX. Toggle Extract tables on to keep table structure as Markdown tables.

Step 3: Download the ZIP with .md + .json

Click Convert. Premium AI parses each page (22-30 seconds for a typical 9-pager). When done, download a ZIP containing:

• document.md — pure Markdown, paste directly into ChatGPT/Claude.
• document_content_list.json — block list with bbox, type, page_no — use for embeddings and RAG chunking.
• images/ — folder with images extracted from the document (if any).

Done! Paste .md into ChatGPT for Q&A, or feed .json into your LangChain/LlamaIndex pipeline.

1. Q&A contracts with ChatGPT/Claude

Upload a contract PDF → convert to Markdown → paste into ChatGPT with: "The following is a contract in Markdown. Summarize Party A's obligations and list any penalty clauses." Markdown preserves headings, so the model understands the clause structure and answers far more accurately than flat PDF paste.

2. Feed academic papers into a RAG pipeline

Research papers contain formulas and tables. LaTeX Markdown lets embedding models (OpenAI text-embedding-3, Cohere) read formulas as text rather than skipping them. JSON bbox enables block-aware chunking that preserves page context.

3. OCR scanned Vietnamese PDFs into an internal knowledge base

Vietnamese scanned documents (lecture notes, official letters, handbooks) can be pushed into Notion, Obsidian, or an internal KB once converted to Markdown. Diacritics preserved at ~99.7% accuracy.

4. Digitize past exams to auto-generate question variants

Teachers can feed past exam papers as Markdown into ChatGPT/Claude to auto-generate question variants, preserving exam structure.

1. For scans: use the original, not photocopies

The AI pipeline works best with 300 DPI or higher scans. Multi-generation photocopies degrade diacritic accuracy.

2. Long files: split first

The 50-page cap is hard. Use Split PDF to break a 200-page file into four 50-page parts, convert each, then concatenate with cat *.md > full.md.

3. Disable "Math formulas" when there are none

Turning on formula extraction for documents with no math (e.g., contracts) can occasionally introduce stray special characters. Off is faster and cleaner.

4. Use JSON bbox for RAG, not Markdown

Markdown is great for pasting into ChatGPT. For production RAG, use .json to chunk by block.type (paragraph/title/table) — retrieval quality beats fixed-length chunking.

5. Verify diacritics

Open .md in VS Code, search for common Vietnamese words to confirm diacritics survived. If they didn't, the source PDF may embed non-standard fonts — try OCR PDF first, then convert.

Error: "File has X pages — exceeds 50-page limit"

The Premium AI pipeline runs on expensive GPU; each job is capped at 50 pages. Fix: use Split PDF to break the file, convert each part, merge Markdown after.

Error: "Premium AI system temporarily disrupted"

The AI engine occasionally restarts to load new models or self-heal. Fix: wait 1-2 minutes and retry. Your job isn't billed — the tool is fully free.

Markdown loses Vietnamese diacritics

Rare (~99.7% accuracy) but possible with very blurry scans. Fix: re-scan the original at 300 DPI, or run OCR PDF first to get a clean text layer before converting.

Table structure broken

Complex tables with merged cells may not export cleanly as Markdown tables. Fix: use the .json bbox (every cell has coordinates) instead of .md, or consider PDF to Excel if you only need the table.

LaTeX formulas misrecognized

Handwritten or low-resolution scans can misread formulas. Fix: review in VS Code with a Markdown+Math extension and manually correct mistakes before feeding to your LLM.

If you're building any AI workflow — document chatbot, RAG knowledge base, summarization pipeline — converting PDF to Markdown is the unavoidable first step. BetaPDF gives you a Premium AI pipeline 100% free, no signup, processing 22-30 seconds per 9 pages.

• ✅ Returns Markdown + JSON bbox in a single ZIP
• ✅ Preserves headings, tables, LaTeX formulas, 99.7% Vietnamese diacritics
• ✅ Ready for ChatGPT, Claude, RAG, embedding pipelines
• ✅ Files auto-deleted after the job completes

Try PDF to Markdown now →

Need more? See the OCR PDF guide for scans, or PDF to Word if you need editable files instead of Markdown.