PDFs are the worst input format for a RAG pipeline. The text is stored as positioned glyphs, not as paragraphs, so a naive extraction gives you broken lines, columns mixed together and tables turned into word soup. Then your chunks are bad, your embeddings are bad, and the model answers from garbage.
Converting each PDF to Markdown first fixes most of this: headings give you natural chunk borders, tables stay readable, and the text is clean enough to embed.
Here are the four problems that hurt retrieval the most, how to check for each, and two ways to do the conversion.
Papers and journals use two columns. A simple extractor reads straight across the page, so line 1 of the left column is glued to line 1 of the right column.
Check: open the output of a two-column paper and read the second paragraph. If the sentences stop making sense halfway through a line, the columns are mixed.
PDF lines end where the page ends, not where the sentence ends. You get infor- on one line and mation on the next, and a keyword search for "information" misses it.
Check: search the output for -\n. A good converter joins the lines of a paragraph and repairs the split words.
A table extracted as plain text loses its columns: the model can no longer tell which number belongs to which header. Markdown tables keep that link and LLMs read them well.
Check: find a table in the PDF and look for | rows in the output.
A scanned page has no text layer at all. The extractor returns an empty string and that page silently disappears from your index. This is the most dangerous case, because nothing fails.
Check: compare the number of pages with the number of characters. A 20-page document with 300 characters is a scan and needs OCR.
For a handful of files with a text layer, this is enough:
import pymupdf4llm # pip install pymupdf4llm
md = pymupdf4llm.to_markdown("paper.pdf")
open("paper.md", "w", encoding="utf-8").write(md)
You will need to add OCR yourself (Tesseract) for scanned pages, and a queue if you have thousands of files.
Disclosure: this one is ours. We publish a hosted converter, PDF to Markdown for LLM & RAG, on Apify. It handles the four problems above: two-column reading order, re-joined paragraphs, Markdown tables, and OCR (Tesseract, 7 languages) for pages without a text layer. You give it URLs and it processes them in parallel.
from apify_client import ApifyClient # pip install apify-client
client = ApifyClient("YOUR_APIFY_TOKEN")
run = client.actor("extas_corp/pdf-to-markdown").call(run_input={
"pdfUrls": [{"url": "https://arxiv.org/pdf/1706.03762"}],
"extractTables": True,
"ocrScannedPages": True,
})
for doc in client.dataset(run["defaultDatasetId"]).iterate_items():
print(doc["title"], doc["pageCount"], "pages,", doc["ocrPages"], "OCR pages")
markdown = doc["markdown"]
Each PDF comes back as one item with status, title, pageCount, ocrPages, tableCount, charCount and markdown. The ocrPages field answers problem 4 directly: you can see which documents were scans.
It is free during launch: you only pay Apify's own platform usage, which the free monthly credits cover for small jobs. Limits: 100 MB per PDF, no password-protected files, and very complex merged-cell tables may come out simplified.
If you only have one file, there is also a free browser version at hackthehydra.com with no account.
Once you have Markdown, split on headings so each chunk is one coherent section:
import re
def chunk_by_heading(markdown, max_chars=4000):
sections = re.split(r"\n(?=#{1,3} )", markdown)
for section in sections:
for i in range(0, len(section), max_chars):
yield section[i:i + max_chars]
Keep the heading inside the chunk: it gives the embedding the context that the body text often lacks.
What breaks most often in your PDF pipeline? I am collecting hard cases to test against.