jgrep
Semantic code search and diff linting that asks Jev whether source chunks match a behavior described in plain language.
Added to Jevfast
How it uses Jev
The CLI chunks source files, batches 16 chunks with 16 Noul questions in a Jev request, and ranks file:line hits by returned probabilities. Its row-scoring mode also passes user-defined Noul, Choice, and Score questions verbatim to Jev.
What Jev decides
Row scoring exampleExample answers · not a recorded Jev response · Source ↗
Question 1 · beauty
YOUR APP
INSTRUCTION
Is beauty the main content of this account?
STATE
One creator record from the input CSV or JSONL row; the README says each row becomes one state. The concrete row values vary by caller.
JEV · NOUL
YesNo
Question 2 · category
YOUR APP
INSTRUCTION
Dominant sub-category?
STATE
One creator record from the input CSV or JSONL row; the README says each row becomes one state. The concrete row values vary by caller.
JEV · CHOICE
- skincare
- makeup
- other
Question 3 · fit
YOUR APP
INSTRUCTION
Fit for a Korean skincare seeding campaign?
STATE
One creator record from the input CSV or JSONL row; the README says each row becomes one state. The concrete row values vary by caller.
JEV · SCORE
0
App workflow
- Choose a search or row-scoring task
The CLI receives a behavior description and code path, a diff, or a CSV/JSONL table with optional typed question JSON.
- Batch Jev judgments
For source search it chunks files and packs 16 chunks with 16 Noul questions per request; row mode sends each record as state with caller-defined typed questions.
- Return bounded results
The CLI ranks source file:line hits by probabilities or writes typed answer columns to a scored table; test selection uses the same source-matching path.
Why it is interesting
It uses parallel typed judgments as an index-free semantic search over source code, then applies the same method to changed-file linting and affected-test selection.