Milvus tests Jev across three agentic-search tasks
Milvus reports integrating Jev, TypeSafe AI’s System One model, into DeepSearcher, MemSearch, and Vector Graph RAG, then comparing it with models or ranking methods already used in those retrieval workflows. The tests cover stopping iterative search, reranking coding-agent memories, and filtering relationships in multi-hop retrieval.
For search stopping, a test on 100 multi-hop questions found both Jev 1.13.0 and the existing DeepSeek Flash service at 93.25% Recall@5. Jev’s median API response latency was 0.55 seconds versus 2.23 seconds, and the post estimates its decision-step cost at about 7.5 times lower. The article says these figures cover the stopping decision, not retrieval, intermediate answer generation, or other calls; Recall@5 measures retrieved evidence, not final-answer accuracy.
Results were mixed elsewhere. On 2,172 MemSearch questions, Jev raised Recall@5 from 74.71% to 79.41%, while Voyage rerank-3 reached 81.87% and had lower estimated cost. In graph retrieval, Jev surpassed GPT-4o-mini on HotpotQA and MuSiQue comparisons as described, but scored 68.87% on MuSiQue against GPT-5-mini’s 73.00%. Milvus characterizes some latency figures as illustrative estimates, not a controlled benchmark.