Authors: Ramani Teegala
Abstract: By December 2021, financial institutions were operating transaction platforms whose end to end behavior increasingly resembled distributed journeys rather than single system events. A single customer initiated action, such as a card purchase, an account to account transfer, or a cross border remittance, could traverse channels, risk engines, limits services, payment rails, settlement systems, dispute workflows, and compliance controls across both internal and external counterparties. This fragmentation created persistent challenges in observability, auditability, and root cause analysis because the underlying data was split across event logs, relational ledgers, message queues, fraud features, and case management systems, each with different identifiers and retention policies. Knowledge graphs matured as a practical representation for integrating heterogeneous entities and relationships, enabling banks to model accounts, customers, devices, merchants, authorizations, postings, reversals, chargebacks, and compliance decisions as a coherent linked structure. In parallel, vector similarity search and embedding based retrieval became increasingly accessible due to open source libraries and emerging vector store implementations, providing a complementary mechanism for approximate matching over high dimensional representations of transactions, sequences, and behavioral signatures. This paper examines how knowledge graphs and vector stores can be combined to represent and analyze financial transaction journeys as understood and practicable by December 2021. The analysis frames the problem through regulated banking constraints, including PCI DSS requirements for cardholder data protection, GLBA expectations for safeguarding customer information, SOX oriented control evidence, Basel Committee guidance on operational risk, and FFIEC style expectations for resilient operations and audit readiness. The paper proposes a conceptual model in which a graph centric system of record captures identity resolution and explicit relationships, while a vector retrieval layer supports similarity based enrichment, anomaly surfacing, and candidate linking for incomplete or ambiguous journey traces. It evaluates architectural trade offs related to consistency, latency, governance, and explainability, emphasizing that approximate methods must be bounded by deterministic controls when outcomes influence fraud actions, customer impact, or regulatory reporting.
DOI: https://doi.org/10.5281/zenodo.19100103
