Convert a PDF to Markdown for RAG
Turn a PDF into structured Markdown and token-sized chunks ready for embedding, using PyMuPDF locally with no GPU or API key.
This recipe turns a PDF into clean Markdown and token-sized chunks you can embed in a vector store. Everything runs locally with PyMuPDF. There is no GPU, no cloud OCR, and no API key.
1. Install PyMuPDF
PyMuPDF 2.x includes the 4LLM Markdown and chunking features.
%pip install -q --disable-pip-version-check pymupdf2. Create a sample PDF
To keep the recipe self-contained, we generate a small two-page PDF with a heading, body text and a table. Swap in your own file by changing pdf_path.
import pymupdf
doc = pymupdf.open()
# Page 1: heading, paragraph and a ruled table
page = doc.new_page()
page.insert_htmlbox(
pymupdf.Rect(50, 50, 545, 200),
"<h1>Quarterly Report</h1>"
"<p>Revenue grew <b>18%</b> year over year, driven by strong demand for document AI.</p>"
"<h2>Regional results</h2>",
)
rows = [
["Region", "Revenue ($M)", "Growth"],
["EMEA", "42.1", "+21%"],
["Americas", "58.7", "+15%"],
["APAC", "23.4", "+19%"],
]
for r, row in enumerate(rows):
for c, value in enumerate(row):
cell = pymupdf.Rect(50 + c * 160, 210 + r * 24, 210 + c * 160, 234 + r * 24)
page.draw_rect(cell, color=(0, 0, 0), width=0.8)
page.insert_text(cell.bl + (6, -8), value, fontname="hebo" if r == 0 else "helv", fontsize=11)
# Page 2: heading and a bullet list
page = doc.new_page()
page.insert_htmlbox(
pymupdf.Rect(50, 50, 545, 300),
"<h2>Outlook</h2>"
"<p>We expect continued growth as more teams build retrieval-augmented generation pipelines.</p>"
"<ul><li>Expand partner network</li><li>Launch new SDK features</li></ul>",
)
pdf_path = "sample.pdf"
doc.save(pdf_path)
print(f"Saved {pdf_path} with {doc.page_count} pages")Saved sample.pdf with 2 pages
3. Preview the first page
Render a page to an image to see what we are working with.
from IPython.display import Image
page = pymupdf.open(pdf_path)[0]
pix = page.get_pixmap(dpi=72, clip=pymupdf.Rect(30, 30, 545, 330))
Image(pix.tobytes("png"))
4. Convert to Markdown
to_markdown() keeps the document structure: headings, bold text, lists and tables. Note: use_ocr is True by default, so we pass use_ocr=False as our sample is born digital.
doc = pymupdf.open(pdf_path)
md = doc.to_markdown(use_ocr=False)
print(md)# **Quarterly Report** Revenue grew **18%** year over year, driven by strong demand for document AI. ## **Regional results** |**Region**|**Revenue ($M)**|**Growth**| |---|---|---| |EMEA|42.1|+21%| |Americas|58.7|+15%| |APAC|23.4|+19%| ## **Outlook** We expect continued growth as more teams build retrieval-augmented generation pipelines. - Expand partner network - Launch new SDK features
5. Chunk for retrieval
to_chunks() splits the document into token-limited chunks that respect headings and tables. Each chunk carries metadata such as page numbers and section path. There are also helpers like to_langchain_documents() and to_llama_nodes().
chunks = doc.to_chunks(max_tokens=60, use_ocr=False)
print(f"{len(chunks)} chunks\n")
for chunk in chunks:
meta = chunk.metadata
print(f"[{chunk.id}] page {meta.page_start} · {' > '.join(meta.section_path)} · {meta.token_count} tokens")
print(chunk.text, end="\n\n")3 chunks [c0] page 1 · Quarterly Report · 24 tokens # **Quarterly Report** Revenue grew **18%** year over year, driven by strong demand for document AI. [c1] page 1 · Quarterly Report > Regional results · 32 tokens ## **Regional results** |**Region**|**Revenue ($M)**|**Growth**| |---|---|---| |EMEA|42.1|+21%| |Americas|58.7|+15%| |APAC|23.4|+19%| [c2] page 2 · Quarterly Report > Outlook · 37 tokens ## **Outlook** We expect continued growth as more teams build retrieval-augmented generation pipelines. - Expand partner network - Launch new SDK features
Next steps
- Pass the chunks to your embedding model and vector store of choice.
- Read the PyMuPDF docs - Chunking for more.
Discuss This Article with the Community
Have a question, a different approach, or something you built after reading this? Share it on the forum or join the Discord, we'd love to hear from you.
