Video unavailable. Open the source below.
NewJevbox
Permission-aware document-tree retrieval with Jev branch routing and passage scoring.
Added to Jevfast
How it uses Jev
Jevbox uses Jev to assess document evidence during search. For each candidate passage, it sends the candidate passage as state and asks how useful that evidence is for answering the whole question, using a Score from 0 to 3. The search can use those evidence scores to identify passages that support an answer; the source also implements Jev Choice routing over document-tree branches.
What Jev decides
Passage usefulness scoring exampleExample answers · not a recorded Jev response · Source ↗
Question 1 · usefulness
YOUR APP
INSTRUCTION
How useful is this passage for answering the entire question, including all requested relationships, entities, time periods, and qualifiers?
STATE
The state is the candidate passage. The user's query is instructions.question; instructions.task supplies the usefulness task and four rubric levels distinguish unrelated, same-topic, partial, and complete evidence.
JEV · SCORE
0
App workflow
- Route through the document tree
For a query, Jev can assign Choice distributions to direct-child document branches. Beam search uses those distributions to prioritize routes while inspecting the tree.
- Score candidate evidence
Jev scores each candidate passage against the query. Its criteria distinguish unrelated material, same-topic material, partial support, and evidence containing the facts needed for the full answer.
- Use relevant passages
The retrieval code uses evidence scores and route results to select material for answering the user's document question.
Why it is interesting
Jevbox applies typed decisions at two levels of document retrieval: Choice can guide which branch to inspect, while Score can judge whether a passage contains the facts and qualifiers the query needs. Its score criteria explicitly distinguish partial evidence from evidence that answers the full question.