LivingRAG: Augmenting Graph RAG with Experience
LivingRAG keeps verified reasoning from earlier queries and feeds it back into later Graph RAG retrieval and generation.
The paper adds a writable “experience” store to a graph-based RAG system. Those stored signals can point retrieval toward entities and passages that helped on related questions, while stored summaries guide the answer pattern. The authors say online QA streams show reuse opportunities through shared entities, graph neighborhoods, and question templates. In multi-hop QA benchmarks, LivingRAG improved accuracy over strong RAG baselines and used fewer completion tokens when prior experience was relevant. ArXiv · AI/CL/LG's note
The paper adds a writable “experience” store to a graph-based RAG system. Those stored signals can point retrieval toward entities and passages that helped on related questions, while stored summaries guide the answer pattern. The authors say online QA streams show reuse opportunities through shared entities, graph neighborhoods, and question templates. In multi-hop QA benchmarks, LivingRAG improved accuracy over strong RAG baselines and used fewer completion tokens when prior experience was relevant. ArXiv · AI/CL/LG's note
score 5