Beyond Fuzzy Matching: A Dual-Augmentation RAG System for Robust Product Reconciliation in Accounting

Accurate product-to-catalog invoice matching is a foundational internal control critical to financial oversight and audit quality, yet it is often bottlenecked by inconsistent vendor descriptions. Traditional rule-based matching fails to address this "long tail" of supplier heterogeneity, leading to costly manual reconciliation. This study presents an end-to-end system for automated invoice reconciliation. We introduce a novel “augment-both-sides” strategy: catalog entries are proactively enriched with LLM-generated keywords and synonyms before vectorization, while incoming invoice line items undergo query expansion to bridge the semantic gap between vendor terminology and master data. A final LLM-based reranker applies context-aware judgment to produce highly accurate Top-3 match candidates. We evaluate this system using three diverse entity resolution benchmark datasets, Abt-Buy, Amazon-Google and Walmart-Amazon, structured to simulate real-world ERP environments. The system achieves a Top-3 Recall of 93.14% to 97.96% across all domains, effectively narrowing the search space for accounting and auditing professionals from thousands of SKUs to a precise set of candidates. These results demonstrate that the architecture functions as a highly reliable intelligent decision aid, standardizing complex reconciliations, and structuring the reconciliation task for subsequent human verification.