HugRAG: Hierarchical Causal Knowledge Graph Design for RAG

Retrieval augmented generation (RAG) has enhanced large language models by enabling access to external knowledge, with graph-based RAG emerging as a powerful paradigm for structured retrieval and reasoning. However, existing graph-based methods often over-rely on surface-level node matching and lack explicit causal modeling, leading to unfaithful or spurious answers. Prior attempts to incorporate causality are typically limited to local or single-document contexts and also suffer from information isolation that arises from modular graph structures, which hinders scalability and cross-module causal reasoning. To address these challenges, we propose HugRAG, a framework that rethinks knowledge organization for graph-based RAG through causal gating across hierarchical modules. HugRAG explicitly models causal relationships to suppress spurious correlations while enabling scalable reasoning over large-scale knowledge graphs. Extensive experiments demonstrate that HugRAG consistently outperforms competitive graph-based RAG baselines across multiple datasets and evaluation metrics. Our work establishes a principled foundation for structured, scalable, and causally grounded RAG systems.

💡 Research Summary

Retrieval‑augmented generation (RAG) has become a cornerstone for extending large language models (LLMs) with up‑to‑date external knowledge. While early RAG pipelines relied on flat text chunking and semantic similarity search, recent graph‑based RAG approaches (e.g., GraphRAG, agentic search, GNN‑guided refinement) introduce structure to improve relevance and interpretability. However, existing graph‑centric methods suffer from two fundamental shortcomings. First, as knowledge graphs grow, intrinsic modularity (dense communities) leads to global information isolation: retrieval often remains trapped inside a local module and fails to reach relevant facts that lie in distant modules. Second, most retrieval policies prioritize semantic proximity, which introduces local spurious noise—nodes that are topically similar but causally irrelevant—thereby degrading precision and increasing hallucination risk. Moreover, standard QA benchmarks focus on short entity answers and do not stress holistic reasoning, masking these issues.

HugRAG addresses both problems through a novel combination of hierarchical graph organization and causal gating. In an offline preprocessing stage, the raw entity‑relation graph extracted from a corpus is partitioned into multiple hierarchical layers using the Leiden community‑detection algorithm. Each module at every level receives a natural‑language summary, forming a multi‑scale backbone H = {H₀,…,H_L}. Crucially, HugRAG then constructs causal gates between module pairs: an LLM‑based estimator scores the plausibility of a causal direction (m_i → m_j); if the score exceeds a threshold τ, a directed gate is added to the set G_c. These gates act as logical bridges that allow traversal across otherwise isolated modules.

During online inference, a query q first selects top‑K seeds at each hierarchy level based on semantic similarity. A gated traversal expands these seeds up to h hops, but traversal is constrained to follow only the causal gates, thereby deliberately breaking modular isolation while avoiding irrelevant edges. The resulting raw subgraph S_raw is then passed through a causal filter implemented by an LLM that discards spurious nodes (V_sp) and retains only those participating in verified causal paths. The filtered subgraph S* is finally fed to the LLM generator, which produces the answer y together with a concise provenance graph.

Extensive experiments were conducted on five domains (power‑grid incidents, medical case studies, legal reasoning, software debugging, and general knowledge) and on a newly introduced cross‑domain benchmark called HolisQA, which emphasizes multi‑step causal reasoning and open‑ended answers. HugRAG consistently outperformed strong baselines—including GraphRAG, LightRA‑G, LeanRA‑G, and CausalRA‑G—by 9‑12 percentage points in recall, 7‑10 points in precision, and up to 16 points in F1. On HolisQA, HugRAG achieved a causal‑accuracy of 84 % compared to the next best 68 %. Ablation studies revealed that removing causal gates drastically reduced recall (‑12 %p) and that using a single‑level seed selection hurt precision (‑9 %p), confirming the complementary role of hierarchy and gating.

The paper also analyses scalability: because only module summaries and a sparse set of causal gates need to be stored in memory, HugRAG can handle graphs with billions of nodes without prohibitive overhead. The explicit causal paths provide intrinsic explainability—users can inspect the filtered subgraph to understand “why” a particular answer was generated, addressing a key limitation of black‑box LLMs.

In conclusion, HugRAG introduces a principled framework that rethinks knowledge organization for graph‑based RAG. By integrating hierarchical modularity with directed causal gates, it simultaneously mitigates global recall gaps and local precision gaps, delivering more faithful, interpretable, and scalable retrieval‑augmented generation. Future work will explore reinforcement‑learning‑driven gate learning, multimodal graph extensions, and incremental updates to the hierarchical structure.