A Markdown-and-Milvus memory system with an optional Jev reranker that independently checks whether each retrieved passage contains the information a query needs.
The query is shared state and each complete passage becomes a Noul question; exact answer coverage and finite probabilities are validated before stable sorting.
Adds an explicit evidence-matching stage after retrieval and raises on failed evaluation rather than silently pretending to rerank.
Fixed-version documentation and implementation material were inspected on 2026-10-24. No memory dataset or live scoring was evaluated. A high passage-match probability is not proof the passage itself is true.
Rerank retrieved memory passages by whether they contain the information the query needs.
Material analysed：
The implementation preserves candidate content without truncation and sends it to TypeSafe. This improves inspectability but means private retrieved memory crosses a remote boundary.
Inspect the supporting source ↗evaluate raises on failed, malformed or incomplete responses instead of returning the original order as if Jev had reranked it. Callers need their own explicit fallback.
Inspect the supporting source ↗Choose it when candidate text may be sent remotely and you want evidence-focused reranking. Enforce a suitable candidate budget before calling it on large private memory stores.
Inspect build_request, scores and rerank in jev_reranker.py. Confirm exact document text, expected answer keys, finite-value checks and stable tie order; compare an empty input with a missing-key request.