HELIX: Hybrid Evidence-Linked Invoice Header Extraction : A Validation-Aware System for Accounting Workflows

Sammanfattning

Invoice automation is viable only when extracted values can be checked before they enter accounting workflows. A wrong due date, value-added tax (VAT) amount, supplier identifier, or payment reference may not only reduce extraction accuracy, but also create financial and review risk. This thesis designs and evaluates HELIX, a hybrid system for header-level invoice information extraction from varied company invoices, including native Portable Document Format (PDF) files and documents that require optical character recognition (OCR). The system combines text recovery, block-based candidate generation, field-specific parsers, optional local large language model (LLM) interpretation, grounding checks, validation rules, and resolver logic. HELIX was evaluated on 110 manually labeled invoice documents and 14 header fields, including invoice numbers, dates, parties, totals, VAT values, payment references, and supplier identifiers. After field-specific semantic recalculation and manual checking of empty reference fields, the evaluated run achieved a positive-field semantic F1 score of 95.4% and a schema-level field accuracy of 94.9%. Invoice date, invoice number, order number when present, currency, and due date performed especially well on labeled positive fields, while the remaining errors were concentrated in party-name role assignment, competing monetary values, and similar business identifiers. Within the evaluated scope, the run supports a bounded proof of concept for checkable invoice header extraction. Larger datasets, further parameter tuning, component ablations, and line-item evaluation are still needed before stronger claims about deployment can be made.