JevGraph
Turn documents into evidence-backed candidate knowledge graphs with typed relation decisions.
Added to Jevfast
How it uses Jev
JevGraph constructs one Choice question per candidate entity relation. State contains the document ID and each candidate's source, target, and evidence; instructions ask Jev to choose the directed relation explicitly supported by that evidence. The selected label and probabilities become recorded edge decisions for later graph review/export.
What Jev decides
Documented relation request exampleExample answers · not a recorded Jev response · Source ↗
Question 1 · edge 1
YOUR APP
INSTRUCTION
Choose a relation supported by the evidence.
STATE
The source test supplies the evidence text Mira founded Northstar. This is a source-defined test example, not a captured model response.
JEV · CHOICE
- founded
- none
App workflow
- Prepare graph candidates
Extract entities and propose directed candidate edges from document evidence using the selected ontology.
- Build typed requests
Group candidates into bounded batches and create one ontology-constrained Choice question for each candidate.
- Evaluate with Jev
Send candidate source, target, evidence, and question criteria to the Jev provider; retain selected labels, probabilities, and request receipts.
- Review and export
Use decisions in the evidence-backed graph workflow and export proposed edges.
Why it is interesting
It turns extraction candidates into evidence-bounded relation decisions and includes explicit `none` and `insufficient_evidence` choices in the ontology, allowing the graph builder to abstain when an edge is unsupported or ambiguous.