HC-RAG: Evidence-Centric Retrieval-Augmented Generation over Heterogeneous Financial Filings
HC-RAG introduces a hierarchical cross-modal RAG framework that organizes financial filings into a typed evidence graph with documents, sections, text units, table units, and metadata nodes The system retrieves evidence through structured document-section-unit paths and aligns textual and tabular evidence in a shared retrieval space Evidence is routed according to four semantic intents: calculation, trend, fact, and comparison, enabling intent-aware text-table fusion The authors introduce Multi-
Analysis
TL;DR
- HC-RAG introduces a hierarchical cross-modal RAG framework that organizes financial filings into a typed evidence graph with documents, sections, text units, table units, and metadata nodes
- The system retrieves evidence through structured document-section-unit paths and aligns textual and tabular evidence in a shared retrieval space
- Evidence is routed according to four semantic intents: calculation, trend, fact, and comparison, enabling intent-aware text-table fusion
- The authors introduce Multi-Doc-2025, a benchmark of 2,327 expert-verified QA pairs from 179 SEC 10-K filings across 87 S&P 500 companies (fiscal years 2022–2024)
- HC-RAG outperforms RAPTOR by 6.6 F1 points on DocFinQA and GraphRAG by 10.9 F1 points on Multi-Doc-2025, with gains driven by accurate section localization, table grounding, cross-document aggregation, and intent-aware routing
Why It Matters
This work addresses a critical gap in financial QA systems: the failure of existing RAG approaches to respect the typed, hierarchical structure of SEC filings and to reason over both text and tables in a query-intent-aware manner. For AI practitioners building enterprise-grade financial analysis tools, HC-RAG demonstrates that structured evidence organization and cross-modal alignment are essential for reliable, verifiable answers over complex regulatory documents.
Technical Details
- Evidence Graph Structure: HC-RAG models filings as a typed financial evidence graph containing five node types—documents, sections, text units, table units, and metadata—enabling path-based retrieval through document-section-unit hierarchies rather than flat chunking.
- Cross-Modal Retrieval: Textual and tabular evidence are aligned in a shared embedding space, allowing the system to jointly retrieve and fuse heterogeneous evidence types relevant to a given query.
- Intent-Aware Routing: Four semantic intent categories (calculation, trend, fact, comparison) guide how retrieved text and table evidence are combined, replacing fixed fusion strategies with query-dependent routing.
- Multi-Doc-2025 Benchmark: 2,327 expert-verified QA pairs drawn from 179 SEC 10-K filings of 87 S&P 500 companies across fiscal years 2022–2024, annotated with intent labels, difficulty levels, and structural evidence attributes.
- Evaluation: HC-RAG achieves +6.6 F1 over RAPTOR on DocFinQA and +10.9 F1 over GraphRAG on Multi-Doc-2025; ablation studies isolate improvements to section localization, table grounding, cross-document evidence aggregation, and intent-aware routing.
Industry Insight
- Financial institutions and regtech companies should prioritize structured, graph-based retrieval over naive chunking when building QA systems over annual reports, as the hierarchical organization of SEC filings is a key signal for accuracy.
- Intent-aware cross-modal fusion is a practical design principle: routing strategies conditioned on query type (e.g., calculation vs. fact) can significantly outperform one-size-fits-all text-table merging in domain-specific RAG pipelines.
- The Multi-Doc-2025 benchmark fills a pressing need for rigorous, multi-document financial QA evaluation and should serve as a standard for comparing future systems in this space.
Disclaimer: The above content is generated by AI and is for reference only.