Hierarchical Reranker wins 2nd in ACM-ICAIF with 0.79 NDCG on financial RAG

Financial analysis demands parsing dense 10-K filings, tabular disclosures, and macroeconomic reports—a task that overwhelms standard retrieval-augmented generation (RAG) systems, which choke on mixed text-table layouts and enormous document volumes. Joohyun Lee and Sungwoo Hong address this with Hierarchical Reranker, a domain-optimized RAG pipeline targeting large-scale financial datasets. The framework integrates three innovations: pre-retrieval optimization (query normalization, keyword expansion, table transformation), a two-stage hierarchical reranking architecture for retrieval precision, and adaptive long-context partitioning and fusion to preserve reasoning accuracy.

Across FinQA, FinanceBench, and ConvFinQA, the system delivered a 0.7918 NDCG@20 score with superior factual consistency, validating its robustness. It also earned second place in the ACM-ICAIF '24 FinanceRAG Challenge. Accepted at FinLLM @ IJCAI-ECAI 2026, the work points toward automated audit reporting and quantitative investment analysis. The source code will be publicly released on GitHub upon acceptance, making the pipeline immediately deployable for generative document analytics.