DEEP-PolyU/LinearRAG

[ICLR 2026] LinearRAG: Linear Graph Retrieval Augmented Generation on Large-scale Corpora

What it solves

LinearRAG 解决了传统 GraphRAG 系统的高成本和慢速问题。具体而言,它消除了在图构建阶段通常需要使用 LLM 来提取实体之间关系的昂贵代币消耗。

How it works

LinearRAG 并不构建显式的关系图,而是采用“无关系”的构建方式。它结合轻量级的实体识别和语义链接,创建图结构。这使系统能够在单次检索过程中通过语义桥接执行多跳推理和深度检索,保持线性时间和空间复杂度。

Who it’s for

该项目面向在大规模语料库上使用 Retrieval‑Augmented Generation(RAG)的开发者和研究者,他们需要图形检索的推理能力,却不想承担 LLM 驱动图构建的高额计算成本。

Highlights

  • Zero LLM token cost: Eliminates LLM usage during the graph building process.
  • Linear Scalability: Operates with linear time and space complexity, making it faster and easier to scale to large datasets.
  • Multi-hop Reasoning: Supports complex reasoning across different pieces of information without needing explicit relational graphs.
  • Context-Preserving: Uses semantic linking to ensure comprehensive contextual understanding.