Abstract

Financial question answering requires retrieval-augmented generation (RAG) systems to remain accurate under stale disclosures, uneven source credibility, and conflicting evidence. Conventional RAG pipelines optimize primarily for semantic relevance, providing limited control over whether retrieved passages are temporally valid, trustworthy, or mutually consistent. We propose TERA-RAG, a reliability-governed multi-agent retrieval architecture that integrates temporal query conditioning, source-reliability-aware evidence scoring, and debate-based conflict resolution in a unified pipeline, filtering and weighting evidence before producing a reliability-weighted consensus answer with auditable reasoning traces. We evaluate TERA-RAG on adversarial subsets of the Fin-RATE benchmark covering disclosure retrieval, evidence conflict, and long-tail temporal consistency. Across 4,500 queries, TERA-RAG achieves 52.2% average exact factual accuracy, outperforming five baselines spanning reliability-aware, corrective-retrieval, self-reflective, and debate-based designs by margins of +15.8 to +27.8 pp. Ablation results confirm that temporal conditioning, reliability weighting, and debate address complementary failure modes. These findings suggest that robust financial RAG requires explicit coordination of temporal validity, source credibility, and epistemic conflict rather than semantic retrieval or debate alone.

Recommended citation: H. N. Truong, T. L. T. Nguyen, Q. V. H. Nguyen, T. P. T. Tran, M. A. Hoang and T. K. Dang, TERA-RAG: Temporal and Epistemic Reliability for Multi-Agent Debate, 2026 23rd International Joint Conference on Computer Science and Software Engineering (JCSSE), Bangkok, Thailand, 2026, pp. 237-242, doi: 10.1109/JCSSE68839.2026.11596782.
Download Paper