Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Breaking the Static Graph: Context-Aware Traversal for Robust Retrieval-Augmented Generation

Kwun Hang Lau, Fangyuan Zhang, Boyu Ruan, Yingli Zhou, Qintian Guo, Ruiyuan Zhang, Xiaofang Zhou

TL;DR

CatRAG tackles the Static Graph Fallacy in HippoRAG 2 by introducing a triad of context-aware mechanisms—Symbolic Anchoring, Query-Aware Dynamic Edge Weighting, and Key-Fact Passage Weight Enhancement—that adapt the knowledge-graph traversal to the user query. By grounding the walk with weak symbolic seeds, pruning irrelevant paths via adaptive edge weighting, and biasing toward evidentiary passages, CatRAG reduces semantic drift and hub bias, improving full chain retrieval and reasoning completeness in multi-hop tasks. Across MuSiQue, 2WikiMultiHopQA, HotpotQA, and HoVer, CatRAG yields consistent recall gains and substantial improvements in FCR and JSR, indicating more grounded, complete evidence chains even when partial recall was already strong. The approach demonstrates that query-conditioned graph steering can bridge partial context and fully grounded multi-hop reasoning, with trade-offs in computation due to on-the-fly LLM scoring and the need for careful hyperparameter tuning.

Abstract

Recent advances in Retrieval-Augmented Generation (RAG) have shifted from simple vector similarity to structure-aware approaches like HippoRAG, which leverage Knowledge Graphs (KGs) and Personalized PageRank (PPR) to capture multi-hop dependencies. However, these methods suffer from a "Static Graph Fallacy": they rely on fixed transition probabilities determined during indexing. This rigidity ignores the query-dependent nature of edge relevance, causing semantic drift where random walks are diverted into high-degree "hub" nodes before reaching critical downstream evidence. Consequently, models often achieve high partial recall but fail to retrieve the complete evidence chain required for multi-hop queries. To address this, we propose CatRAG, Context-Aware Traversal for robust RAG, a framework that builds on the HippoRAG 2 architecture and transforms the static KG into a query-adaptive navigation structure. We introduce a multi-faceted framework to steer the random walk: (1) Symbolic Anchoring, which injects weak entity constraints to regularize the random walk; (2) Query-Aware Dynamic Edge Weighting, which dynamically modulates graph structure, to prune irrelevant paths while amplifying those aligned with the query's intent; and (3) Key-Fact Passage Weight Enhancement, a cost-efficient bias that structurally anchors the random walk to likely evidence. Experiments across four multi-hop benchmarks demonstrate that CatRAG consistently outperforms state of the art baselines. Our analysis reveals that while standard Recall metrics show modest gains, CatRAG achieves substantial improvements in reasoning completeness, the capacity to recover the entire evidence path without gaps. These results reveal that our approach effectively bridges the gap between retrieving partial context and enabling fully grounded reasoning. Resources are available at https://github.com/kwunhang/CatRAG.

Breaking the Static Graph: Context-Aware Traversal for Robust Retrieval-Augmented Generation

TL;DR

CatRAG tackles the Static Graph Fallacy in HippoRAG 2 by introducing a triad of context-aware mechanisms—Symbolic Anchoring, Query-Aware Dynamic Edge Weighting, and Key-Fact Passage Weight Enhancement—that adapt the knowledge-graph traversal to the user query. By grounding the walk with weak symbolic seeds, pruning irrelevant paths via adaptive edge weighting, and biasing toward evidentiary passages, CatRAG reduces semantic drift and hub bias, improving full chain retrieval and reasoning completeness in multi-hop tasks. Across MuSiQue, 2WikiMultiHopQA, HotpotQA, and HoVer, CatRAG yields consistent recall gains and substantial improvements in FCR and JSR, indicating more grounded, complete evidence chains even when partial recall was already strong. The approach demonstrates that query-conditioned graph steering can bridge partial context and fully grounded multi-hop reasoning, with trade-offs in computation due to on-the-fly LLM scoring and the need for careful hyperparameter tuning.

Abstract

Recent advances in Retrieval-Augmented Generation (RAG) have shifted from simple vector similarity to structure-aware approaches like HippoRAG, which leverage Knowledge Graphs (KGs) and Personalized PageRank (PPR) to capture multi-hop dependencies. However, these methods suffer from a "Static Graph Fallacy": they rely on fixed transition probabilities determined during indexing. This rigidity ignores the query-dependent nature of edge relevance, causing semantic drift where random walks are diverted into high-degree "hub" nodes before reaching critical downstream evidence. Consequently, models often achieve high partial recall but fail to retrieve the complete evidence chain required for multi-hop queries. To address this, we propose CatRAG, Context-Aware Traversal for robust RAG, a framework that builds on the HippoRAG 2 architecture and transforms the static KG into a query-adaptive navigation structure. We introduce a multi-faceted framework to steer the random walk: (1) Symbolic Anchoring, which injects weak entity constraints to regularize the random walk; (2) Query-Aware Dynamic Edge Weighting, which dynamically modulates graph structure, to prune irrelevant paths while amplifying those aligned with the query's intent; and (3) Key-Fact Passage Weight Enhancement, a cost-efficient bias that structurally anchors the random walk to likely evidence. Experiments across four multi-hop benchmarks demonstrate that CatRAG consistently outperforms state of the art baselines. Our analysis reveals that while standard Recall metrics show modest gains, CatRAG achieves substantial improvements in reasoning completeness, the capacity to recover the entire evidence path without gaps. These results reveal that our approach effectively bridges the gap between retrieving partial context and enabling fully grounded reasoning. Resources are available at https://github.com/kwunhang/CatRAG.
Paper Structure (35 sections, 5 equations, 2 figures, 9 tables)

This paper contains 35 sections, 5 equations, 2 figures, 9 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Dense Retriever
  4. Structure-Aware RAG
  5. Dynamic & Adaptive Retrieval
  6. Methodology
  7. Preliminaries
  8. Symbolic Anchoring
  9. Query-Aware Dynamic Edge Weighting
  10. Adaptive Entity Contextualization
  11. Stage I: Coarse-Grained Candidate Pruning
  12. Stage II: Fine-Grained Semantic Probability Alignment
  13. Key-Fact Passage Weight Enhancement
  14. Unified Retrieval Process
  15. Experimental Setup
  16. ...and 20 more sections

Figures (2)

  • Figure 1: Comparison of graph traversal between HippoRAG 2 and CatRAG. We illustrate the retrieval process for the multi-hop query "Which university did Marie Curie's doctoral advisor attend?". In HippoRAG 2 (top), the static graph structure causes semantic drift; probability mass is diverted to high-weight generic edges (e.g., Marie Curie$\rightarrow$Radioactivity), missing the downstream evidence ENS. CatRAG (bottom) prevents this by applying (1) Symbolic Anchoring, injecting "University" as a weak seed, (2) Query-Aware Dynamic Edge Weighting amplifying relevant paths (e.g., Attend in ENS) while pruning irrelevant ones, and (3) Key-Fact Passage Weight Enhancement to strength, boosting relevant context edge. This steers the random walk to successfully retrieve the complete evidence chain for ENS.
  • Figure 2: Distribution of PPR-Weighted Node Strength ($\mathcal{S}_{ppr}$). Comparison of the HippoRAG 2 versus CatRAG. The distribution for CatRAG is shifted to the left, indicating a reduction in the retrieval of high-degree "Hub" nodes. The dashed lines represent the mean $\mathcal{S}_{ppr}$ for each method.