GOSU: Retrieval-Augmented Generation with Global-Level Optimized Semantic Unit-Centric Framework

Xuecheng Zou, Ke Liu, Bingbing Wang, Huafei Deng, Li Zhang, Yu TangarXiv 2025 · DOI · arXiv

Introduction

Standard Retrieval-Augmented Generation (RAG) systems typically rely on fixed-length text chunks, which often leads to a "local view" problem. Relationships between entities are frequently fragmented across chunk boundaries, and coreference resolution—identifying that different terms in separate chunks refer to the same entity—becomes a significant challenge. While graph-based RAG methods attempt to address this by building entity-relationship graphs, they often fall into two traps: focusing solely on simple binary relations (missing complex event context) or using overly coarse hypergraphs that lose fine-grained detail. GOSU (Global-Level Optimized Semantic Unit-Centric Framework) addresses these limitations by shifting the focus to globally optimized "Semantic Units" (SUs).

Methodology

GOSU introduces a three-stage pipeline designed to capture both fine-grained binary relationships and complex, multi-entity events through a unified structure.

The overall framework of GOSU, illustrating the three-stage process: Global-level Semantic Unit Optimization, SU-Centric Knowledge Graph Construction, and Dual-Phase Retrieval and Generation.

1. Global-Level Semantic Unit Optimization

Instead of extracting facts from isolated chunks, GOSU identifies "candidate" SUs across the entire corpus. An LLM is then used to judge similarity, cluster related units, and deduplicate them globally. This ensures that a specific event or concept is represented consistently, regardless of its distribution across the source documents.

2. SU-Centric Knowledge Graph Construction

The knowledge graph is built by centering it on these refined SUs. Fine-grained entities and binary relations are extracted conditioned on the SUs. This creates a balanced "hub-and-spoke" architecture where the SU represents a multi-entity event (n-ary relation) while preserving specific links between individual entities.

3. Dual-Phase Retrieval and Generation

Retrieval utilizes a three-tier keyword strategy: low-level (specific entities), mid-level (semantic unit phrases), and high-level (broad themes). Keyword-matched subgraphs are combined with "Semantic Unit Completion" to provide the LLM with a coherent, contextually rich prompt for final response generation.

Experimental Evaluation

The framework was evaluated using the UltraDomain Benchmark (covering Agriculture, CS, Law, and Mixed domains) and a specialized Hypertension Guidelines dataset for high-stakes medical scenarios. Evaluation was conducted using an LLM-as-a-judge protocol to rank GOSU against baselines including NaiveRAG, LightRAG, HiRAG, and HyperGraphRAG.

Comparative performance of GOSU against various RAG baselines across multiple domains and evaluation dimensions.

Key Results

  • Significant Performance Gains: GOSU achieved an average win-rate margin of over +70% against NaiveRAG across all dimensions (Comprehensiveness, Diversity, Empowerment, and Overall).
  • Superiority over Graph Baselines: Even compared to advanced systems like HyperGraphRAG, GOSU maintained a substantial lead, notably achieving an +88.2% gap in Comprehensiveness.
  • Ablation Insights: Removing the Global Optimization (GO) module led to sharp performance declines, particularly in complex domains like Law and Hypertension, confirming that global consistency is vital for high-quality retrieval.

Conclusion and Future Work

GOSU demonstrates that disambiguating semantic units globally before graph construction effectively solves the fragmentation issues inherent in local-extraction methods. While the framework currently focuses on text, future work aims to incorporate multimodal data (images and tables) and improve the computational efficiency of the pairwise similarity matching used during the global optimization phase.

Summary updated

BibTeX
@misc{zou2025gosu,
  author = {Xuecheng Zou and Ke Liu and Bingbing Wang and Huafei Deng and Li Zhang and Yu Tang},
  title = {{GOSU: Retrieval-Augmented Generation with Global-Level Optimized Semantic Unit-Centric Framework}},
  year = {2025},
  eprint = {2509.00449},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2509.00449},
}
Figure