NewVector Graph RAG
Use Jev to rank graph relations and source passages before a language model answers a multi-hop question.
Added to Jevfast
NewHow it uses Jev
Jev scores candidate graph relations against the user's question and shared graph context. The highest-ranked relations seed source passages. A second set of Noul questions ranks up to 16 graph and dense-retrieval excerpts before the answer model uses the selected evidence.
What Jev decides
Rank candidate relationsExample answers · not a recorded Jev response · Source ↗
Question 1 · r{relation id}
YOUR APP
INSTRUCTION
Does this relation supply a needed fact or intermediate multi-hop link?
STATE
The user question and candidate_graph containing labeled relation IDs and text. One dynamic r{relation_id} question is generated per relation.
JEV · NOUL
YesNo
Rank source passagesExample answers · not a recorded Jev response · Source ↗
Question 1 · d{i}
YOUR APP
INSTRUCTION
Does this target excerpt contain useful supported evidence for the question?
STATE
The user question and up to 16 excerpts keyed d0, d1 and so on.
JEV · NOUL
YesNo
App workflow
- Retrieve and expand
Milvus vector retrieval finds entities and relations, then expands the candidate subgraph.
- Rank relations
Jev Noul scores order the candidates; the top 64 relation seeds identify source passages.
- Rank passages
Alternate graph passages with up to 10 dense candidates, deduplicate to at most 16, and score each excerpt with Jev.
- Answer
Send the highest-ranked passages to the answer generator and return retrieval details.
Why it is interesting
Two bounded decision stages connect multi-hop graph facts with supporting source text. Each excerpt must support a useful fact itself; shared context alone does not make it relevant.