PyMuPDF4LLM 2.1 changes the packaging: the namesake distribution is a compatibility placeholder, while the 4LLM code ships in PyMuPDF 2.x. Installing pymupdf4llm==2.1 resolved pymupdf==2.1 in this test, and the pymupdf4llm import remained usable. A synthetic text PDF retained its sentence; a textless fixture returned whitespace with OCR disabled.
To compare output from the same document, see MarkItDown's acceptance checks and Docling's table checks, then inspect text, columns and reading order separately.
Check the implementation package
The upstream repository README describes the move and recommends installing pymupdf. The example pins both distributions to 2.1 so a later dependency resolution cannot silently select another implementation. Older descriptions of package size or dependencies should not be carried forward without checking the version.
Create a text and a textless control
Use pymupdf.open() to make a one-page PDF with Dispatch NOTE-31 leaves at 08:30. at (72,72). Make a second PDF containing only a rectangle. That is a control for absent text, not a scanned photograph or a representative scan benchmark.
These commands use a POSIX shell. The test ran in Linux; Windows shell behavior was not tested.
python -m pip install "pymupdf==2.1" "pymupdf4llm==2.1"
python - <<'PY'
import pymupdf4llm
for name in ["text.pdf", "no-text.pdf"]:
text = pymupdf4llm.to_markdown(name, use_ocr=False)
print(name, repr(text), bool(text.strip()))
PY
Inspect output with OCR explicitly disabled
to_markdown(..., use_ocr=False) preserved the full sentence in the first PDF and returned only whitespace for the rectangle-only PDF. Both input hashes stayed unchanged. A pipeline should check stripped output before sending it to search or RAG. An empty result is different from a crash and does not justify inventing missing text.
Untested behavior and selection
OCR, multicolumn layout, tables, page chunking, and image handling are documented from official v2.1 material and were not locally tested. No accuracy or performance comparison was conducted. For Word or HTML table inputs, the existing Docling guide covers a different tested scope. Choose based on the source text layer and structure you need, then verify a known-answer sample.
Official sources and test scope
Codex ran this synthetic-data test in a credential-free, non-root, network-disabled container. It is not a claim that the site owner tested it. Package version: 2.1; documentation snapshot: 1200f2aea929.