PDF Processing

ADX uses PyMuPDF (fitz) to parse PDFs into structured DocumentGraph objects. This guide covers PDF-specific features.

What Gets Extracted

• Text blocks with bounding boxes and classification (heading, paragraph, header, footer)
• Tables with row/column structure and cell values
• Sections inferred from heading hierarchy and font sizes
• Figures/images detected with bounding boxes
• Page-level metadata — dimensions, text density

Heading Detection

Text blocks are classified by font size relative to the page average:

• Font size > 1.5x average → heading
• Position in top 10% of page → header
• Position in bottom 10% of page → footer
• Everything else → paragraph

Headings drive section detection. Sections are nested by font size to build a document outline.

Table Extraction

PyMuPDF's find_tables() extracts tabular data with cell boundaries. Tables are converted to structured Table objects with:

• Row and column counts
• Header row detection
• Cell-level values
• Bounding boxes for citation
• Markdown rendering via table.to_markdown()

Scanned PDF Detection

ADX warns when a page appears to be scanned (image-based) rather than text-based. Pages with fewer than 20 characters of extracted text trigger a possibly_scanned warning.

For scanned PDFs, consider using an OCR pre-processing step before uploading to ADX.

Citations

Every text block and table gets a Citation with:

• page — 1-indexed page number
• bbox — [x0, y0, x1, y1] coordinates in PDF points
• text — the source text snippet

Citations flow through to extractions, so every extracted field traces back to its source location in the PDF.

Example

from adx import ADX

dn = ADX()
doc_id = dn.upload("contract.pdf")

# Get document structure
structure = dn.structure(doc_id)
for section in structure["sections"]:
    print(f"Section: {section['title']} (pages {section['start_page']}-{section['end_page']})")

# Read a specific page
page = dn.get_page(doc_id, page_number=1)
for block in page["text_blocks"]:
    print(f"[{block['type']}] {block['text'][:80]}...")

# Get a table
tables = structure["tables"]
if tables:
    table = dn.get_table(doc_id, tables[0]["id"])
    print(table["markdown"])

Limitations

• Scanned PDFs require OCR pre-processing (not built in)
• Complex multi-column layouts may produce fragmented text blocks
• Form fields (AcroForm) are not yet extracted
• Annotations and comments are not yet extracted