DEEP-PolyU/LinearRAG

What it solves

LinearRAG addresses the high cost and slow speed of traditional GraphRAG systems. Conventional methods often rely on LLMs to extract relations between entities to build a knowledge graph, which consumes significant tokens and time. LinearRAG provides a way to build these graphs without needing LLM-based relation extraction, making the process faster and more scalable for large datasets.

How it works

Instead of explicit relational graphs, LinearRAG uses a "relation-free" construction method. It employs lightweight entity recognition and semantic linking to create a graph structure. By using semantic bridging, the system can perform multi-hop reasoning and deep retrieval in a single pass, allowing it to navigate complex information without the overhead of traditional graph construction.

Who it’s for

This project is for developers and researchers building RAG systems who need to handle large-scale corpora where the cost and latency of LLM-driven graph construction are prohibitive.

Highlights

• Zero LLM token cost for construction: Eliminates the need for LLMs to identify relations during the graph-building phase.
• Linear complexity: Operates with linear time and space complexity, ensuring high scalability.
• Multi-hop reasoning: Supports deep retrieval and complex reasoning via semantic bridging without explicit relational graphs.
• Context-preserving: Uses semantic linking and entity recognition to maintain comprehensive contextual understanding.