Check Python and PyTorch first, download models into a known folder, then convert the practice PDFs. For a few short quotes, choose manual entry when extraction setup is more work than the task.
In this guide
Will extraction remove the part you keep repeating?
Docling is useful to evaluate when source documents repeatedly need a working text representation before you can review the terms. With a few short quotes, manual entry may be easier to maintain. With supplier CSVs, begin from the structured data and preserve each original.
Test one representative document before processing a folder. Inspect a price, a quantity, and a condition in their original positions. Then decide whether extraction plus correction is useful for the formats you actually receive.
Selectable text or scanned pages?
Check both text and tables visually. A scan can require OCR; the right options depend on your installed pipeline. A readable export still needs checking for shifted rows, missing symbols, and detached conditions.
Local processing or a remote option?
Docling documents local models and separately enabled remote-service options. Check the configuration you use, allow for initial model downloads, and review any later hosted AI step separately.
What do you need before the first conversion?
Choose a supported 64-bit Python installation and a matching PyTorch build. This walkthrough uses Python 3.12 as an explicit environment choice, not Docling's only supported version. Use a new folder with the three practice PDFs and the Python script.
Windows: can the environment import PyTorch?
Install Python first and confirm that the launcher finds the chosen version. These PowerShell commands use the environment interpreter directly, avoiding activation-script policy changes. Run the import check before waiting for model downloads.
py -3.12 --version py -3.12 -m venv .venv .\.venv\Scripts\python.exe -m pip install --upgrade pip .\.venv\Scripts\python.exe -m pip install docling .\.venv\Scripts\python.exe -m pip check .\.venv\Scripts\python.exe -c "import torch; print(torch.__version__)"
Download models separately from the conversion
The default pipeline downloads several model assets on first use. A progress bar for one table model is not the total installation size. Allow space for both packages and model caches; the selected pipeline and versions determine which files are needed. Inspect their reported sizes instead of assuming every setup has the same download.
The advanced-options guide documents prefetching and an artifacts path. Inspect the installed download tool's help, download the needed models, and keep that directory for repeat runs.
.\.venv\Scripts\docling-tools.exe models download --help .\.venv\Scripts\docling-tools.exe models download --output-dir .\models .\.venv\Scripts\python.exe extract-practice-quotes.py --artifacts-path .\models .\.venv\Scripts\python.exe -m pip freeze > environment.txt
Which Windows stage failed?
Hugging Face settings vary by client version. Current documentation describes HF_HUB_DISABLE_SYMLINKS=1 for copy-based caching. Version 0.36 describes Windows fallback and Developer Mode instead. HF_HUB_DISABLE_SYMLINKS_WARNING only hides a warning; it does not grant link privileges or repair another package's own link operation. Check which library raised the exception before adopting a workaround.
| Observed failure | What to check | Next step |
|---|---|---|
| c10.dll / WinError 1114 during import | Python architecture, PyTorch build, runtime dependencies | Use the platform instructions; rerun import first |
| WinError 1314 while caching | Package and path attempting a symbolic link | Follow version-matched cache guidance or authorized OS settings |
| Download stalls | Network, free disk, model file, partial download | Retry the download stage; keep useful cached files |
| Extracted values shift or disappear | Original PDF and extraction layout | Correct from source or enter the short quote manually |
| TableModel04_rs model file missing | Incomplete table-model artifacts, even though imports succeeded | Finish model prefetch and point the script at that directory |
macOS and Linux: use the matching interpreter
Read the architecture notes, especially Intel Mac and accelerator-specific packages. Save environment.txt after a conversion you have checked. An arbitrary version pin should not be presented as a known-good setup.
python3 -m venv .venv .venv/bin/python -m pip install docling .venv/bin/python -c "import torch; print(torch.__version__)" .venv/bin/python extract-practice-quotes.py
What counts as a successful first run?
- The script exits successfully and creates one nonempty Markdown file for each PDF.
- Amount, currency, quantity, and conditions match their locations in the originals.
- A rerun stops before overwriting an existing output.
- environment.txt records the successful environment; an import alone is not a completed conversion.
References
Documentation checked October 20, 2026.
Use it in a task
Supplier Quote Comparison for Business Owners: Extract, Check, Compare
Continue the task
Inspect the extracted fields and the later context-selection stage before trusting a document-based answer.