GRACE: Graph-Guided Repository-Aware Code Completion through Hierarchical Code Fusion

2025-09-07

AI Summary

To address the limitations of context scarcity and inadequate modeling of cross-file semantic/structural dependencies in repository-level code completion, this paper proposes a graph-guided retrieval-augmented generation (RAG) framework. Unlike conventional text-similarity-based RAG approaches, our method constructs a multi-granularity code graph integrating file structure, abstract syntax trees (ASTs), call graphs, and data-flow graphs. We design a graph neural network–driven hybrid retriever with topology-aware attention for re-ranking. Furthermore, we introduce a hierarchical structured context fusion mechanism that explicitly preserves function call chains and class inheritance relationships. Evaluated on multiple repository-level benchmarks using DeepSeek-V3, our approach achieves substantial improvements over state-of-the-art methods: an 8.19% absolute gain in Exact Match (EM) and a 7.51% gain in Edit Similarity (ES).

Technology Category

Natural Language Processing: Code Generation / Program Synthesis from Natural LanguageMachine Learning: Graph-based Machine LearningData Mining & Knowledge Management: Linked Open Data, Knowledge Graphs & KB Completion

Application Category

Search and Retrieval-Augmented AI: Retrieval-Augmented Generation (RAG) and multi-modal RAGGraph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphsSemantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs

Abstract

LLMs excel in localized code completion but struggle with repository-level tasks due to limited context windows and complex semantic and structural dependencies across codebases. While Retrieval-Augmented Generation (RAG) mitigates context scarcity by retrieving relevant code snippets, current approaches face significant limitations. They overly rely on textual similarity for retrieval, neglecting structural relationships such as call chains and inheritance hierarchies, and lose critical structural information by naively concatenating retrieved snippets into text sequences for LLM input. To address these shortcomings, GRACE constructs a multi-level, multi-semantic code graph that unifies file structures, abstract syntax trees, function call graphs, class hierarchies, and data flow graphs to capture both static and dynamic code semantics. For retrieval, GRACE employs a Hybrid Graph Retriever that integrates graph neural network-based structural similarity with textual retrieval, refined by a graph attention network-based re-ranker to prioritize topologically relevant subgraphs. To enhance context, GRACE introduces a structural fusion mechanism that merges retrieved subgraphs with the local code context and preserves essential dependencies like function calls and inheritance. Extensive experiments on public repository-level benchmarks demonstrate that GRACE significantly outperforms state-of-the-art methods across all metrics. Using DeepSeek-V3 as the backbone LLM, GRACE surpasses the strongest graph-based RAG baselines by 8.19% EM and 7.51% ES points on every dataset. The code is available at https://anonymous.4open.science/r/grace_icse-C3D5.

Problem

• Addressing repository-level code completion challenges in LLMs
• Overcoming structural relationship neglect in code retrieval methods
• Preserving code dependencies lost in naive snippet concatenation

Innovation

• Multi-level code graph unifies static and dynamic semantics
• Hybrid Graph Retriever integrates structural and textual similarity
• Structural fusion mechanism preserves essential code dependencies