HiKEY: Hierarchical Multimodal Retrieval for Open-Domain Document Question Answering

Researchers introduce HiKEY, a hierarchical multimodal retrieval framework designed to improve document-based question answering systems by leveraging document structure as a core retrieval signal. The system addresses critical limitations in existing approaches by implementing a coarse-to-fine retrieval strategy and demonstrating significant performance improvements on ODQA benchmarks.

HiKEY represents a meaningful advancement in retrieval-augmented generation technology, tackling two fundamental problems that plague current document question-answering systems: inefficient document routing and fragmented multimodal evidence integration. Rather than treating documents as flat collections of text chunks, the framework reconstructs logical hierarchies through Document Hierarchical Parsing, enabling more precise navigation through large document corpora. This structural approach mirrors how humans naturally process complex documents, moving from broad categories to specific sections.

The broader context reveals growing recognition within AI research that document structure contains valuable signal often discarded by simpler chunking approaches. As enterprises deploy QA systems over massive document repositories—financial filings, technical manuals, legal contracts—the ability to quickly locate correct documents and synthesize information across tables, figures, and text becomes increasingly critical. Existing page-level and chunk-based methods struggle with scale and multimodal integration, creating bottlenecks in practical industrial deployments.

For developers building enterprise search and knowledge management systems, HiKEY offers tangible improvements: up to 12.9% better retrieval recall and 6.8% end-to-end performance gains translate to more accurate answers and reduced hallucination. This matters particularly for regulated industries where answer accuracy directly impacts compliance and decision-making quality. Organizations using RAG pipelines could potentially reduce computational costs through better initial document pruning.

The token-efficient evidence packing strategy addresses a persistent constraint in LLM-based systems, enabling more complete context windows without proportional cost increases. Looking ahead, the integration of hierarchical reasoning into production retrieval systems could become standard practice, particularly as document complexity and corpus size continue increasing across enterprise applications.

• →HiKEY improves retrieval recall by up to 12.9% and end-to-end QA performance by 6.8% through hierarchical document structure awareness
• →The framework uses Document Hierarchical Parsing to explicitly encode parent-child relationships, treating document structure as a first-class retrieval signal
• →A coarse-to-fine retrieval strategy enables rapid search space pruning while maintaining multimodal evidence integration across tables, figures, and text
• →Token-efficient evidence packing reduces computational requirements by strategically selecting the most discriminative information within budget constraints
• →Performance gains suggest hierarchical approaches may become standard in enterprise document QA systems handling large-scale industrial corpora