Query-Aware Subgraph Packing: A Knapsack Optimization Paradigm for Graph Retrieval-Augmented Generation

Chenjie Qiu, yongqi luo, Bingfeng chen, Boyan Xu, Ruichu Cai, Zhifeng Hao

Keywords: Natural Language Processing, Graph Based Learning, Information Retrieval

Abstract: Graph Retrieval‑Augmented Generation (GraphRAG) has recently emerged as a task paradigm for injecting graph‑structured knowledge into large language models (LLMs), yet most existing approaches still rely on flat, similarity‑based retrieval that ignores topology and uses static encoders, producing redundant or structurally incoherent evidence. In this paper, we propose GraphPack, a query‑aware GraphRAG framework that overcomes these limitations by casting subgraph selection as a 0–1 knapsack optimization. For every natural language query, GraphPack packs the most informative subgraph under a size budget by jointly maximizing semantic relevance and minimizing structural redundancy. The selected subgraph is then encoded by a query‑aware graph encoder whose parameters are conditioned on the query, allowing node representations to adapt dynamically to user intent. Extensive experiments on multiple knowledge-intensive graph benchmarks demonstrate that GraphPack achieves state-of-the-art performance, showcasing its strong capability in addressing structural and contextual challenges under supervised learning, cross-domain settings, and zero-shot scenarios.

Primary Area: Deep learning (e.g., architectures, generative models, optimization for deep networks, foundation models, LLMs)

Submission Number: 27887