Zilliz's open-source Vector Graph RAG integrates the Jev decision model with knowledge graph retrieval, achieving 86.61% Recall@5 on MuSiQue and 2Wiki datasets while reducing model call latency, outperforming LLM-based reranking methods in both accuracy and speed.
The article examines FlowRAG, a new variant of GraphRAG that shifts retrieval from similarity‑based text chunk ranking to constructing explicit, frequency‑aware reasoning paths, detailing its three‑step design, benchmark improvements, efficiency gains, and ablation results that reveal how it mitigates entity sparsity and noise propagation.
The article presents the MDER‑DR two‑stage framework that tackles semantic loss in knowledge‑graph triple indexing by generating context‑aware entity summaries and using an LLM‑driven decompose‑parse retrieval loop, achieving up to 66% performance gains on multi‑hop question answering benchmarks.
The paper introduces DLLM‑Searcher, which equips diffusion large language models with a two‑stage training pipeline and a P‑ReAct inference scheme, allowing the model to issue tool calls while simultaneously reasoning, yielding 14‑22% end‑to‑end speedup and matching or surpassing traditional autoregressive agents on multi‑hop QA benchmarks.
The Search‑R2 framework integrates error detection, localization, and correction into a reinforcement‑learning loop for search‑enhanced reasoning, achieving notably larger accuracy gains on difficult multi‑hop QA tasks than baseline methods, even when those baselines receive higher sampling budgets.
Relink introduces a ‘reason‑and‑construct’ paradigm that builds knowledge‑graph paths during inference, combining a high‑precision factual graph with a high‑recall potential‑relation pool, using query‑driven dynamic path expansion and contrastive alignment to markedly improve multi‑hop QA performance and robustness to sparse knowledge.
The article explains why traditional retrieval‑augmented generation fails in multi‑hop scenarios due to overly large chunks, introduces SentGraph’s sentence‑level graph that trims retrieval units and encodes logical relations, details offline construction and online inference steps, and shows experimental gains and remaining limitations.
