pip install belegradar

Turn a PDF invoice into validated structured data: vendor, line items, VAT breakdown per rate (MwSt-Aufschluesselung), totals, checksum-validated IBAN, and a paragraph 14 UStG required-fields check, with a flag on anything that does not add up.

The design rule: the LLM reads, plain Python verifies. The model copies values verbatim; deterministic code does all locale number parsing (1.234,56 vs 1,234.56), date parsing, and arithmetic checks.

Quickstart

from belegradar import extract_text, extract_invoice, validate

text = extract_text(open("invoice.pdf", "rb").read())  # raises on scans
invoice, tokens = extract_invoice(text)                # needs GROQ_API_KEY
report = validate(invoice)

print(invoice.vendor_name, invoice.gross_total, invoice.currency)
for flag in report.flags:
    print(flag.severity, flag.field, flag.message)
print("Pflichtangaben:", report.pflichtangaben)

What it validates (no LLM involved)

• gross_total == net_total + sum(VAT amounts), within rounding tolerance
• line items sum to net_total; each quantity * unit_price == total
• each VAT bucket: rate% of base == amount
• IBAN ISO 13616 mod-97 checksum, pure Python
• VAT id country formats (DE/AT/CH/NL/FR/GB, extendable)
• paragraph 14 UStG Pflichtangaben: presence of every field a German invoice must carry to be valid for input-tax deduction

Utilities

from belegradar import iban_is_valid, vat_id_is_valid
iban_is_valid("DE89 3704 0044 0532 0130 00")  # True (spaces tolerated)
vat_id_is_valid("ATU12345678")                # True

Honest constraints

• Text-layer PDFs only. Scanned or photographed invoices raise NoTextLayerError. There is no reliable free OCR at zero infrastructure cost, and pretending would produce silently wrong data. OCR is a roadmap item.
• Extraction is not perfect; the product stance is extraction with validation and flags, which is what real back-office workflows need.
• The LLM provider is swappable (Groq by default via GROQ_API_KEY); it is never trusted with arithmetic or locale conversion.

Used in production by

The live demo at beleg.nagysolution.com runs this package. Accuracy is measured by an eval suite over a gallery of fake invoices; results are published in the repository.

License

MIT. Built by Nandor Nagy as part of a public portfolio.