Milvus Jev Reranker
Rank candidate documents by Jev evidence-relevance judgments for a supplied claim.
Added to Jevfast
How it uses Jev
`JevRerankFunction` sends the query as a scientific claim and creates one Noul question per candidate document asking whether it provides evidence relevant to that claim. It reads each answer's yes-probability, sorts documents by that score, and returns up to `top_k` results.
What Jev decides
Scientific claim evidence reranking (illustrative)Example answers · not a recorded Jev response · Source ↗
Question 1 · d0
YOUR APP
INSTRUCTION
Given the document text, decide whether it provides evidence relevant to the claim.
STATE
A claim is paired with a list of candidate documents; each document receives a generated key `d0`, `d1`, and so on. The displayed example describes one representative candidate.
JEV · NOUL
YesNo
App workflow
- Configure reranker
Provide an API key directly or through `TYPESAFE_API_KEY` and select a model name.
- Create per-document questions
Build one Noul question for each document and put the query in state as a scientific claim.
- Call Jev
Send all candidate questions in one HTTP request to the System One endpoint.
- Rank results
Read each Noul probability, sort documents in descending order, and return at most `top_k` results.
Why it is interesting
It adapts a yes/no evidence judgment into a familiar reranking interface: each document is assessed against the query, then the returned probability becomes the ranking score.