Skip to content

Camelot and Tabula PDF table loaders

These optional adapters extract tables from PDFs and return the same page dictionaries as other ExtractThinker loaders. Each page contains content (PDF text), page_number (one-based), and tables (a list of tables, each a list of rows of strings). Pages with no detected tables remain in the result with tables=[]; blank cells and table boundaries are retained.

Camelot

python -m pip install camelot-py

Install camelot-py, not the unrelated package named camelot. See the Camelot API reference for parser options and the dependencies of your installed version.

from extract_thinker import DocumentLoaderCamelot, CamelotConfig

loader = DocumentLoaderCamelot(CamelotConfig(flavor="lattice"))
pages = loader.load("invoice.pdf")
for page in pages:
    for table in page["tables"]:
        print(page["page_number"], table)

lattice is useful for ruled tables; stream uses text spacing. network, hybrid, and auto depend on the installed Camelot version. Additional parser options can be passed through read_pdf_kwargs, for example {"split_text": True}. Source, page range, password and flavor cannot be overridden through that dictionary.

Tabula

python -m pip install tabula-py
java -version

Install a Java runtime and make java available on PATH. Installing the Python package alone does not install Java. See the tabula-py API reference.

from extract_thinker import DocumentLoaderTabula, TabulaConfig

loader = DocumentLoaderTabula(TabulaConfig(
    lattice=True,
    read_pdf_kwargs={"force_subprocess": True},
))
pages = loader.load("invoice.pdf")

Use either lattice=True for ruled tables or stream=True for text layout. guess defaults to True; java_options can configure the runtime. read_pdf_kwargs accepts additional parser options, such as area or columns, but cannot override source, page selection, output format, or explicitly configured settings. The loader requests each page separately because Tabula's table JSON does not consistently carry page identity. This preserves blank pages and multiple tables, with additional invocation overhead on long documents.

Shared configuration and behavior

Both configurations accept cache_ttl=300, password=None, and vision_enabled=False.

from io import BytesIO

stream = BytesIO(pdf_bytes)
pages = loader.load(stream)
loader.set_vision_mode(True)
loader.set_max_image_size(1024)
selected = loader.load_pages("invoice.pdf", [3, 1])

Streams retain their position. Temporary PDFs used by parser backends are deleted after success or failure. Vision mode adds a PNG image per page; PDFium reads text and renders images locally. load_pages selects after parsing and does not reduce table-extraction work. Parser and Java failures propagate; a failed run is not cached as an empty document.

These adapters do not perform OCR. For scans, use an OCR loader or supply a PDF with a usable text layer. Table quality depends on the chosen parser and document layout. The loaders are available on the repository's main branch after release 0.1.14; install from source until a release containing them is published.