DEEP-PolyU/LinearRAG

What it solves

LinearRAG addresses the high cost and slow speed of traditional GraphRAG systems. Specifically, it eliminates the expensive LLM token consumption required during the graph construction phase, which typically involves using an LLM to extract relations between entities.

How it works

Instead of building explicit relational graphs, LinearRAG uses a "relation-free" construction method. It combines lightweight entity recognition and semantic linking to create a graph structure. This allows the system to perform multi-hop reasoning and deep retrieval via semantic bridging in a single retrieval pass, maintaining linear time and space complexity.

Who it’s for

This project is for developers and researchers working with Retrieval-Augmented Generation (RAG) on large-scale corpora who need the reasoning capabilities of graph-based retrieval without the prohibitive computational costs of LLM-driven graph construction.

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.