DocNavRAG: Document-Structured Graph RAG with Stateful Evidence Construction for Complex Document Question Answering

2026-08-02

AI Summary

This work addresses the challenge of integrating evidence across sections and documents in complex question answering. The authors propose an intelligent retrieval framework grounded in a navigable document structure graph, which encodes both hierarchical document organization and cross-regional relationships. A stateful evidence tracking mechanism guides an agent to perform structured operations—such as locating, jumping, expanding, and retrieving—to dynamically accumulate relevant evidence. By moving beyond conventional fixed traversal strategies or weakly structured retrieval paradigms, the approach achieves significant improvements on four long- and multi-document QA benchmarks, yielding average gains of 7.8% in answer quality and 17.7% in context sufficiency.

Abstract

Answering complex questions over large document collections requires assembling complementary evidence across sections and documents. GraphRAG offers structured retrieval but typically uses fixed traversal, while agentic RAG operates over weakly structured interfaces. Our key insight is that agents should navigate document structure within and across documents rather than repeatedly search from scratch. We introduce DocNavRAG, which organizes document hierarchies and cross-region relations into a navigable graph, exposes graph operations for locating, navigating, expanding, and fetching, and maintains an evolving evidence state to guide retrieval until sufficient evidence is collected. Across four long- and multi-document QA benchmarks, DocNavRAG improves answer quality and context sufficiency over the strongest baseline by 7.8\% and 17.7\% on average.

Problem

• complex document question answering
• evidence assembly
• document navigation
• structured retrieval
• multi-document QA

Innovation

• Document-Structured Graph RAG
• Stateful Evidence Construction
• Navigable Graph
• Complex Document QA
• Agent-Based Retrieval