LiteRAG: Cost-Efficient Graph-Based Retrieval-Augmented Generation

TL;DR - LiteRAG is a graph-based RAG method that replaces retrieval-time LLM control with query-conditioned graph exploration and compact reasoning-chain construction. It maintains or improves retrieval quality while substantially reducing latency, cost, and context size.

• Achieves the highest reported overall quality on DistComp (0.798), a multi-hop retrieval benchmark for distributed-systems papers.
• Reduces per-query latency by over 100× and cost by over 99% versus GraphRAG Global and DRIFT.
• Matches LinearRAG’s overall quality on UltraDomain while using roughly 14× fewer tokens.
• Ablations identify query-adaptive thresholding and community-aware hub penalization as the main contributors to token efficiency.