Milvus 实测 Jev 替代 LLM 调用：三类智能体搜索场景表现分化
Milvus 用三个真实搜索工作流实测 Jev 替代 LLM 调用，延迟降 4 倍、成本降 7.5 倍，还给出了哪些场景该换哪些不该换的结论。
Milvus 团队在三个智能体搜索工作负载中测试 Jev 能否替代通用 LLM 调用。在 DeepSearcher 的停止判断任务中，Jev 与 DeepSeek Flash 持平于 93.25% Recall@5，中位延迟从 2.23 秒降到 0.55 秒，成本约低 7.5 倍。在 MemSearch 中，Jev 将 Recall@5 从 74.71% 提升到 79.41%，但仍不及成本更低的 Voyage rerank-3（81.87%）。在 Vector Graph RAG 测试中，Jev 在 HotpotQA 和 MuSiQue 上超过 GPT-4o-mini，HotpotQA 上与 GPT-5-mini 差距在 1 点以内，MuSiQue 上落后 4.13 点。结论是 Jev 适合高频、边界清晰的语义判断，复杂多跳检索仍需更强的生成模型。
We tested whether Jev could replace some of the LLM calls inside three agentic search workloads. 1️⃣ In 𝗗𝗲𝗲𝗽𝗦𝗲𝗮𝗿𝗰𝗵𝗲𝗿, Jev matched DeepSeek Flash at 93.25% Recall @5 . Median latency for the stopping decision fell from 2.23 seconds to 0.55, while estimated cost was about 7.5x lower. For this kind of narrowly defined stopping decision, Jev can replace the general-purpose LLM without sacrificing aggregate retrieval quality. 2️⃣ In 𝗠𝗲𝗺𝗦𝗲𝗮𝗿𝗰𝗵, Jev raised Recall @5 from 74.71% to 79.41%. Voyage rerank-3 still performed better, reaching 81.87% at a lower cost. For this workload, the dedicated reranker remains the better choice: higher retrieval quality at a lower cost. 3️⃣ In 𝗩𝗲𝗰𝘁𝗼𝗿 𝗚𝗿𝗮𝗽𝗵 𝗥𝗔𝗚, Jev beat GPT-4o-mini on HotpotQA and MuSiQue. It came within one point of GPT-5-mini on HotpotQA, but trailed it by 4.13 points on MuSiQue. Jev is a viable option when latency matters and some loss in recall is acceptable, but a stronger generative model remains better for difficult multi-hop retrieval. Overall, Jev worked best for frequent, bounded semantic decisions. Its advantage narrowed when a specialized model already fit the task or when the decision required reasoning across multiple pieces o milvus.io/blog/jev-vs-ll… ll results: https://t.co/7NdCtU1XFu 💬 0 🔄 0 ❤️ 0 👀 49 ⚡
