Back to AI Research

AI Research

LivingRAG: Augmenting Graph RAG with Experience | AI Research

Key Takeaways

  • LivingRAG is a framework designed to improve Graph Retrieval-Augmented Generation (RAG) by allowing systems to learn from previous queries.
  • Graph-based RAG improves multi-hop question answering by organizing evidence as a knowledge graph.
  • However, most existing RAG systems process each query in isolation and discard useful reasoning from the LLM's response after inference.
  • As a result, later related queries need to retrieve evidence and reason from scratch.
  • We propose LivingRAG, a Graph RAG framework with writable and reusable reasoning experience.
Paper AbstractExpand

Graph-based RAG improves multi-hop question answering by organizing evidence as a knowledge graph. However, most existing RAG systems process each query in isolation and discard useful reasoning from the LLM's response after inference. As a result, later related queries need to retrieve evidence and reason from scratch. We propose LivingRAG, a Graph RAG framework with writable and reusable reasoning experience. LivingRAG adds a writable experience store to a graph-based retrieval backbone, enabling verified experiences to be reused during inference in two ways. Stored graph signals help retrieval find entities and passages that were useful in earlier related queries. Stored summaries provide a reference reasoning pattern for answer generation. We analyze online QA streams and find reusable signals from shared entities, graph neighborhoods, and question templates. Experiments on multi-hop QA benchmarks show that LivingRAG improves accuracy over strong RAG baselines and reduces completion-token use when relevant prior experience is reused.

LivingRAG is a framework designed to improve Graph Retrieval-Augmented Generation (RAG) by allowing systems to learn from previous queries. While standard RAG systems process each question in isolation and discard the reasoning used to generate an answer, LivingRAG introduces a "writable experience store." This store captures and reuses verified reasoning patterns and graph-based evidence, helping the system answer subsequent, related questions more accurately and efficiently.

Addressing the "From Scratch" Problem

Most current RAG systems treat every query as a new task, forcing the model to retrieve evidence and perform reasoning from the beginning every time. Researchers Yuzhuo Cui, Zongye Zhang, and Qingjie Liu identified that this approach ignores valuable signals that persist across online question-answering streams. These signals include shared entities, similar graph neighborhoods, and recurring question templates. By failing to store these insights, systems miss opportunities to optimize performance for related queries.

How LivingRAG Works

LivingRAG functions by adding a writable memory component to a graph-based retrieval backbone. The process follows three main steps:

  • Experience Capture: After a query is answered, the system creates a candidate experience containing the query, the retrieved evidence, the reasoning summary, and the final answer.

  • Quality Control: To prevent the accumulation of errors or redundant information, the system uses "quality gates." A candidate is only saved if it is factually grounded in the retrieved evidence (verified via an NLI model) and is sufficiently novel compared to existing entries in the store.

  • Reuse: When a new query arrives, the system retrieves relevant stored experiences. It uses these to guide the graph retriever toward useful evidence neighborhoods and provides the LLM with a "reasoning scaffold"—a compact summary of how a similar problem was previously solved.

Performance and Efficiency

In experiments across multiple multi-hop question-answering benchmarks—including 2Wiki, HotpotQA, MuSiQue, and WixQA—LivingRAG consistently outperformed strong RAG baselines in accuracy. Beyond improving answer quality, the researchers found that reusing prior experience reduced the number of completion tokens required for generation, which can lower API costs. The system achieves these gains without requiring changes to the underlying model parameters or the static graph structure.

Considerations and Limitations

While LivingRAG improves performance, it introduces specific computational overhead. The system must perform a scan of the experience store for every new query to identify relevant prior knowledge. Additionally, the write-back process requires NLI-based grounding checks, which add a computational cost for each candidate experience. The researchers note that the effectiveness of the system depends on the presence of reusable signals in the data stream; if a query is entirely unique and shares no structural or thematic overlap with past questions, the system effectively reverts to its base RAG performance.

Comments (0)

No comments yet

Be the first to share your thoughts!