latex-llm-cleaner 0.6.0

Flatten and clean LaTeX files for LLM consumption

LaTeX LLM Cleaner

A Python CLI tool that takes a LaTeX .tex file or compiled .pdf and produces a cleaned text version optimized for LLM consumption. Combines functionality from tools like flachtex, arxiv_latex_cleaner, and pandoc into a single utility.

Installation

pip install latex-llm-cleaner

# With PDF support:
pip install latex-llm-cleaner[pdf]

Or from source:

pip install .
# or for development:
pip install -e ".[dev]"

Usage

latex-llm-cleaner paper.tex                    # output to stdout
latex-llm-cleaner paper.tex -o cleaned.tex     # output to file
latex-llm-cleaner paper.tex --no-bibliography  # skip bib inlining
latex-llm-cleaner thesis.pdf -o thesis.md      # extract text from PDF
latex-llm-cleaner thesis.pdf --ocr -o thesis.md  # OCR with LaTeX equation recovery

All features are on by default. Disable individual steps with --no-* flags:

latex-llm-cleaner INPUT_FILE [options]

Options:
  -o, --output FILE          Write to FILE (default: stdout)
  --no-flatten               Disable \input/\include flattening
  --no-comments              Disable comment removal
  --no-bibliography          Disable bibliography inlining
  --no-figures               Disable figure summary substitution
  --figure-summary-suffix S  Suffix for summary files (default: _summary.txt)
  --ocr                      Use Surya vision OCR (recovers LaTeX equations, slower)
  --encoding ENC             File encoding (default: utf-8)
  -v, --verbose              Print processing info to stderr

PDF Input

For compiled PDFs (e.g., theses, published papers), latex-llm-cleaner extracts the text as markdown, preserving table structure and dropping images. This requires the optional pdf extra:

pip install latex-llm-cleaner[pdf]
latex-llm-cleaner thesis.pdf -o thesis.md

Tables are output as markdown tables with | delimiters. Images are noted as [picture omitted] placeholders. The .tex pipeline flags (--no-flatten, etc.) are ignored for PDF input since they don't apply.

OCR mode (equation recovery)

The default PDF extraction is fast but loses display equations. For compiled LaTeX PDFs, the --ocr flag uses Surya vision-based OCR to recover equations as LaTeX source:

pip install latex-llm-cleaner[ocr]
latex-llm-cleaner thesis.pdf --ocr -o thesis.md

This reconstructs inline math as $...$ and display equations as $$...$$ with full LaTeX notation. It's slower (~30s/page on Apple Silicon) but dramatically more accurate for math-heavy documents. Requires Python ≤ 3.13.

Processing Pipeline (.tex files)

The four steps run in this order (each operates on the output of the previous step):

1. Flatten includes — inline \input{}, \include{}, and \subfile{} recursively, with cycle detection
2. Remove comments — strip % comments while respecting \% escapes and verbatim environments
3. Inline bibliography — use a pre-compiled .bbl file if available (common in arXiv downloads), otherwise parse .bib files; replaces \bibliography{} with a \begin{thebibliography} block
4. Figure summary substitution — replace \includegraphics with text descriptions when summary files are available

Figure Summaries

LLMs (as of early 2026) are still poor at extracting precise information from complex figures in papers — dense plots, multi-panel layouts, small labels, etc. To work around this, latex-llm-cleaner can replace figures with equivalent text descriptions.

For each image (e.g., figs/plot.png), place a summary file alongside it with the configured suffix:

figs/plot.png              ← the image
figs/plot_summary.txt      ← the text summary

What to put in a summary

A summary should be data-equivalent to the figure: it should convey the same information a reader would get from looking at the figure, and nothing more. Avoid editorial commentary, interpretation, or conclusions that aren't visually present in the figure itself.

Good example:

Bar chart with four groups (A, B, C, D). Method X scores 0.92, 0.87, 0.76, 0.81. Method Y scores 0.85, 0.91, 0.80, 0.74. Error bars show standard deviation across 5 runs.

Bad example:

This figure clearly demonstrates the superiority of Method X, which aligns with our hypothesis.

Accessibility benefits

LaTeX has limited built-in support for producing accessible output — generated PDFs typically lack alt-text for figures, making them difficult to navigate with screen readers. These same summary files can serve as alt-text source material when compiling to tagged PDF or HTML, improving accessibility beyond the LLM use case.

Development

pip install -e ".[dev]"
pytest