Marker PDF to Markdown: convert one file first

Check dependencies before installing

Marker’s current README specifies Python and PyTorch requirements and describes device-dependent inference backends. Read that section for your machine before assuming pip installation is the entire setup. A model download or missing inference server can be the first-run bottleneck rather than the PDF itself.

Convert a single PDF

Install the documented marker-pdf package in an isolated environment, then use the single-file entry point. Select a new output folder so repeated trials do not obscure which run produced a file.

python -m pip install marker-pdf
marker_single sample.pdf --output_format markdown --output_dir marker-output

The command follows the current README. It has not been run as a performance or quality benchmark by NarrowTools; dependencies and model assets must be configured first.

Separate the local run from optional LLM processing

The README describes an optional --use_llm route. Do not enable it by copying a benchmark command when you intend a local-only trial. Provider configuration, document transmission, and usage charges are separate decisions. Record the options used so another run can reproduce your setup.

Compare the output

Use the same small fixture as the Docling guide. Print its HTML to PDF and compare both converters against the reference content. Check the literal SKU, price, table rows, and the paragraph after the table. Use a scan from your own workflow as a second test; a clean generated PDF cannot reveal scan quality.

Review terms before packaging a product

Marker’s repository distinguishes the code license from the model-weight terms. Inspect both for the version and intended usage rather than assuming one license covers everything. Before committing to a service, evaluate model setup, storage, uncertain output, and correction work as well as conversion speed.