PyMuPDF loader¶

This loader is available on main after release 0.1.14. Until the next package release, install from source with the optional parser:

pip install "git+https://github.com/enoch3712/ExtractThinker.git" pymupdf

PyMuPDF is an optional dependency with its own license terms.

from extract_thinker import DocumentLoaderPyMuPDF, PyMuPDFConfig

loader = DocumentLoaderPyMuPDF(PyMuPDFConfig(
    include_bbox=True,
    extract_tables=True,
    vision_enabled=True,
    dpi=150,
))
pages = loader.load("invoice.pdf")

The loader accepts a PDF path or BytesIO stream. Each returned page contains content and a one-based page_number. Enable extract_tables for a tables list, include_bbox for regions, and vision mode for PNG image bytes.

This loader reads the PDF's existing text layer; it does not perform OCR on scanned text. Choose an OCR loader or use vision extraction for image-only PDFs. Native PyMuPDF calls from this adapter are serialized across threads.

Configuration	Default	Meaning
`cache_ttl`	`300`	Cache duration in seconds
`password`	`None`	Password for encrypted PDFs
`extract_tables`	`False`	Include tables detected by PyMuPDF
`include_bbox`	`False`	Include source text blocks and normalized coordinates
`vision_enabled`	`False`	Render a PNG of each page
`dpi`	`150`	Image resolution, a positive integer

loader.set_max_image_size(1200) bounds the rendered image's longest edge. loader.load_pages("invoice.pdf", [3, 1]) selects pages in the requested order. Selection currently loads the document before filtering.

Source coordinates¶

Each regions entry can be parsed as DocumentRegion:

from extract_thinker import DocumentRegion

region = DocumentRegion.model_validate(pages[0]["regions"][0])
print(region.text)
print(region.bounding_box.page, region.bounding_box.x0, region.bounding_box.y0)

Coordinates are fractions between 0 and 1, measured from the top-left of the rendered page, including its rotation. x0, y0, x1, y1 mean left, top, right and bottom. The loader supplies these coordinates; they are not estimates generated by an LLM. Empty pages have an empty regions list.

Extractor preserves these source regions and table metadata in model input. When asking an LLM to associate a field with a source region, verify that association separately; a valid rectangle does not prove a correct extraction.