Open-source semgrep finds code and logs by meaning, not regex, with cross-language matching

udmrzn · x · 2026-09-20

Developer uehaj open-sourced jev-semgrep, a grep that matches lines by meaning instead of regex, built on TypeSafe AI's Jev System One model.

Jev doesn't generate text; it answers typed yes/no questions with probabilities, so each line is scored against a natural-language proposition and filtered by a threshold.
Performance: 30 lines batched per request, 8 requests in parallel, a 210-line file finishes in under a second. Zero dependencies, a single Node.js 20.12+ file, runnable via npx @uehaj/semgrep -e "network failure" app.log.
Meanings combine with AND/OR/NOT (-e, -a, -v) plus grep-like options: -r recursive, -l filenames only, -A/-B/-C context, --color, and --level loose/strict thresholds.
Unlike embedding search, it evaluates propositions rather than topic distance: "customer wants a refund" vs "refund processed" score 0.98 and 0.10 despite being nearly identical in embedding space, so boolean logic works directly.
Language-agnostic: queries and text can each be any language. Mixed 14-line documents in 6 languages were queried in Japanese/English/French/Russian with zero misses and zero false positives.

More from coding & agent