Question 1

Is this vector search with better marketing?

Accepted Answer

No. It asks Jev whether a proposition holds for each line, reading line and query together, so negation and who-did-what change the answer; an embedding of the line is fixed before it ever sees your query. Because the scores are calibrated, one threshold works across queries where cosine similarity needs top-k or per-query tuning.

Question 2

What does a query over a big log cost, and can I index it?

Accepted Answer

There is no index. sys1grep reads the files in front of you and every query pays for the whole corpus again, which the README names as the flip side: for repeated queries over a large fixed corpus a vector index is cheaper and faster. Inside one pass it batches 30 lines per request with 8 requests in flight, finishing a 210-line file in under a second.

Question 3

The package was renamed - what do I install now?

Accepted Answer

From 0.5.0 the npm package is @uehaj/sys1grep, the commands are sys1grep and git sys1grep, the environment prefix is SYS1GREP_ and the config file is ~/.config/sys1grep/.env. The old @uehaj/semgrep and SEMGREP_* names still work for one minor release with a deprecation line, and 1.0.0 drops them.

Question 4

Will a Japanese query work as reliably as an English one?

Accepted Answer

One query does match across Japanese, French, Russian, German, Spanish, Chinese and Korean, but TypeSafe documents English as the most accurate language and the author's own tests show Japanese meanings wobbling more near the threshold. When a query is borderline, phrase the meaning in English.

Question 5

How is sys1grep installed and invoked?

Accepted Answer

It is zero-dependency and a single file, requiring Node.js 20.16+ and fetch. The renamed package is @uehaj/sys1grep with commands `sys1grep` and `git sys1grep`; for example `sys1grep -n -e "customer is angry or frustrated" tickets.txt`.

Question 6

How do regex terms behave?

Accepted Answer

A term shaped like /pattern/flags is matched locally as a plain JavaScript regex with no request at all. It prefilters its AND term, so a cheap regex in front of a meaning cuts both the bill and the wait, and a query of regex terms alone sends nothing.

Question 7

How do AND, OR and NOT work across meanings?

Accepted Answer

Each meaning yields an independent calibrated probability, so logical AND and NOT are plain boolean operations rather than negative-query tricks. One threshold (0.5) works across queries, unlike cosine similarity which needs top-k or per-query tuning.

Jev Semgrep

About this project

Highlights

Questions about this entry

Reactions & coverage

From the Jev guides

From the Jev AI model community