Vector Graph RAG 用 Jev 两阶段决策把 Recall@5 提到 86.61%
Milvus 上的 Vector Graph RAG 用 Jev 先筛关系、再排段落，Jev 1.14 在 MuSiQue 与 2Wiki 上达到 86.61% Recall@5。
We pushed Vector Graph RAG to 𝟴𝟲.𝟲𝟭% 𝗥𝗲𝗰𝗮𝗹𝗹@𝟱 𝗮𝗰𝗿𝗼𝘀𝘀 𝟮,𝟬𝟬𝟬 𝗾𝘂𝗲𝘀𝘁𝗶𝗼𝗻𝘀, ahead of every LLM reranker we tested and onto the quality-latency Pareto frontier. Built on Milvus, Vector Graph RAG follows connections between entities and relations to find evidence across documents. Our biggest improvement came from a 𝘁𝘄𝗼-𝘀𝘁𝗮𝗴𝗲 𝗝𝗲𝘃 𝗱𝗲𝗰𝗶𝘀𝗶𝗼𝗻-𝗺𝗼𝗱𝗲𝗹 𝗽𝗶𝗽𝗲𝗹𝗶𝗻𝗲: 1. 𝗙𝗶𝗹𝘁𝗲𝗿 𝗿𝗲𝗹𝗮𝘁𝗶𝗼𝗻𝘀. Jev selects useful relations from the expanded graph, preserving the connections needed to reach relevant evidence. 2. 𝗥𝗮𝗻𝗸 𝗽𝗮𝘀𝘀𝗮𝗴𝗲𝘀. We fetch those relations’ source passages, merge them with direct vector-search results, and deduplicate. Jev then ranks the passages for the user’s question. With Jev 1.13, the two-stage recipe lifted Recall@5 from 𝟴𝟬.𝟬𝟰% 𝘁𝗼 𝟴𝟲.𝟬𝟳%. Jev 1.14 reached 𝟴𝟲.𝟲𝟭% across MuSiQue and 2Wiki. Useful relations get us to the evidence. Passage ranking decides what makes the final cut. Code and benchmarks: https://github.com/zilliztech/vector-graph-rag