Search with Jev and Milvus

Watch out

The README states up front that these are runnable integration examples and not quality benchmarks, built on small synthetic datasets so you can inspect every candidate and decision, so no accuracy, latency or cost numbers should be expected. The captured text is truncated partway through the scenario table: six of the nine scenarios are described, while the remaining descriptions and the guardrails, curation and evaluation notebooks are not in what was captured. Nine notebooks are listed alongside pyproject.toml, uv.lock and a .gitignore, so dependencies are pinned by uv, but setup instructions, Milvus deployment details, embedding configuration and where credentials go are not documented here. One described scenario depends on a second model: the stopping notebook has Gemini Flash generate evidence-driven queries while Jev gates stopping within three rounds before a grounded answer or an abstention, and the capture does not say whether the other notebooks stay single-model. The README's motivating failures, a result similar to a question without answering it, and a cached answer matching the topic while using the wrong software version, are stated as design rationale without supporting measurement. Copy reviewed against the linked page on 2026-10-04.