Jgrep
Filter lines and records with natural-language predicates and Jev probabilities.
Added to Jevfast
How it uses Jev
jgrep turns each selected input record into a Noul question: does this line or other chosen text unit fit the user's natural-language description? The returned probability is compared with the configured threshold, then matching records are printed, counted, ranked, or emitted in a structured format. The CLI also supports several descriptions in one call per record.
What Jev decides
Check whether a line fits a descriptionExample answers · not a recorded Jev response · Source ↗
Question 1 · d0
YOUR APP
INSTRUCTION
Does this input text fit the user's natural-language description?
STATE
For each input record, state contains the candidate text line, and the Noul instruction is the caller's description. The selected example illustrates the documented 'frustrated user' use case; it is not an observed response.
JEV · NOUL
YesNo
App workflow
- Read and shape input
The CLI reads files, stdin, or structured records and selects lines, paragraphs, whole files, functions, chunks, or diff hunks according to its flags.
- Build judgments
It creates one Noul question per description, includes the selected text as state, and submits records concurrently with a bounded number of calls.
- Filter and order results
Returned probabilities are checked against the threshold; ordered mode emits matches in input order, while unordered mode emits answers as they arrive.
- Format output
The CLI prints matching text or filenames, counts, probability-ranked output, JSON/JSONL/CSV records, or grep-compatible exit status.
Why it is interesting
It makes semantic filtering composable with ordinary Unix pipelines. Input order and bounded concurrency are preserved by default, so a natural-language predicate can sit in a live log or file workflow with familiar grep-style output and controls.