Enhancing Mean Time To Resolution (MTTR) In High-Frequency Financial Platforms: A Dual-Stage Retrieval-Augmented Generation (RAG) Approach With Metadata-Aware Re-Ranking

Abstract

Financial systems operating in high-frequency trading and real-time settlement environments face critical challenges in incident response, where rapid diagnosis directly impacts financial exposure and regulatory compliance. Traditional diagnostic workflows require engineers to manually correlate millions of log entries with proprietary code and documentation under extreme time pressure, resulting in extended resolution cycles. While Large Language Models offer powerful reasoning capabilities, parametric models hallucinate when lacking domain-specific knowledge, and conventional Retrieval-Augmented Generation systems fail to differentiate between data sources of varying epistemic fidelity. This introduces a Dual-Stage RAG architecture with Metadata-Aware Re-Ranking that addresses the knowledge heterogeneity problem by explicitly prioritizing transactional ground-truth logs over general documentation. The architecture implements query parsing to extract transaction identifiers, enabling filtered retrieval of operational data, combined with a weighted re-ranking function that assigns source credibility weights based on evidential hierarchy principles. Experimental validation using the RAGAS framework demonstrates substantial improvements: the Hybrid system achieves a significant reduction in diagnostic latency, improves Context Precision substantially, and achieves strong Faithfulness scores while reducing hallucination rates significantly. The system successfully concentrates high-fidelity transactional logs in top retrieval positions, ensuring LLMs receive better-targeted evidence that enables precise root cause identification with explicit temporal and quantitative evidence citations, offering actionable remediation guidance for production incident response teams.