grep, but the pattern is a question in natural language.
$ cat hn-feed.jsonl | jevgrep --json "is this about a company actually shipping something?"
$ jevgrep --whole "does this module talk to the network?" src/**/*.py
$ jevgrep -c "is this an authentication failure?" /var/log/app.loggrep -cE 'ERROR|WARN' finds seven lines in the sample log. Four of them are actually
authentication failures. A regex can only match spelling; the other three are an HTTP 500, a
connection pool exhausting and a rate limit.
No embeddings, no vector index, no build step. Each record is sent to the Jev evaluation model as a typed yes/no question, which answers with a probability plus its own confidence in about 200 ms for roughly $0.00002. That is cheap enough to run a predicate over a whole log file, and fast enough to sit in the middle of a pipe.
Because Jev can only answer the question it was given, it cannot wander off and invent an explanation. A record either clears the threshold or it does not.
jevutils adds the other two verbs, built on this
package: jevsort ranks records by how strongly a question holds, and jevuniq collapses
records that mean the same thing. They compose in a pipe.
$ jevuniq tickets.txt | jevsort "how urgent is this?" | jevgrep -c "is this an outage?"curl -fsSL https://raw.githubusercontent.com/anup-a/jevgrep/main/install.sh | shOr directly, if you already have one of these:
uv tool install git+https://github.com/anup-a/jevgrep # recommended
pipx install git+https://github.com/anup-a/jevgrepFrom a clone, for hacking on it:
uv sync
uv run jevgrep --helpSet credentials in the environment or in a .env file next to where you run it
(see .env.example):
JEVGREP_API_KEY=... # or TYPESAFE_API_KEY
JEVGREP_BASE_URL=https://ai-gateway.vercel.sh/v4/ai
JEVGREP_MODEL=typesafe-ai/jevjevgrep [OPTIONS] PREDICATE [FILE...]
With no FILE, reads stdin. One record per line by default, exactly like grep.
| Flag | Meaning |
|---|---|
-v, --invert-match |
Select records the predicate is false for |
-c, --count |
Print only the count of matching records |
-l, --files-with-matches |
Print each matching origin once |
-n, --line-number |
Prefix output with the line number |
-q, --quiet |
Print nothing; signal via exit code |
--color WHEN |
auto (default, when stdout is a terminal), always, never |
-t, --threshold FLOAT |
Probability needed to match (default 0.5) |
--min-confidence FLOAT |
Records Jev is less sure than this about never match (default 0.0) |
-j, --jobs INT |
Requests in flight (default 8) |
--json |
Parse each input line as JSON and send the object as the record |
--whole |
One record per file instead of per line |
--explain |
Prefix each result with [p=… c=…] |
--jsonl |
Emit one JSON verdict per line, for piping onward |
--stats |
Report record count, latency and cost on stderr |
Same contract as grep, so && and if work as you expect:
| Code | Meaning |
|---|---|
0 |
At least one record matched |
1 |
No record matched |
2 |
Something went wrong (bad config, unreadable file, API failure) |
These are two different things and the distinction is the reason this tool is worth using.
--thresholdis about the record: how probable is it that the predicate holds?--min-confidenceis about the model: how sure is Jev of its own answer?
A record can be p=0.99, c=0.3: "almost certainly yes, but I am guessing". Raising
--min-confidence pushes those into an uncertain bucket that never matches and is reported
on stderr, so ambiguous records get escalated to a human (or a bigger model) instead of
silently landing on one side.
$ jevgrep --min-confidence 0.7 --explain "is this a security incident?" incidents.log
[p=0.98 c=0.91] 2026-09-19 root login from 203.0.113.9, no MFA
jevgrep: 3 records were too uncertain to classify (see --explain)Under --explain, both numbers are coloured by how much they can be trusted: green at or
above 0.9, amber at or above 0.7, red below. A record the model is guessing at is obvious
at a glance. Colour follows grep's rules: on when stdout is a terminal, off when piped,
NO_COLOR respected, --color=always to force it. --jsonl output is never coloured.
# Filter a JSONL feed on a field-aware predicate, keep the original lines
cat jobs.jsonl | jevgrep --json "is this role remote and paid in USD?" > remote.jsonl
# Which source files reach the network?
jevgrep --whole -l "does this module open a socket or make an HTTP request?" src/**/*.py
# Structured output for a downstream tool
jevgrep --jsonl "is this a stack trace?" app.log | jq -r 'select(.probability > 0.9) | .text'
# Use it as a test, grep-style
jevgrep -q "does this diff delete a test?" <(git diff) && echo "heads up"Two things bite here, and jevgrep handles both explicitly rather than telling you to disable verification:
httpxpins its own certifi bundle and ignoresSSL_CERT_FILE. jevgrep readsJEVGREP_CA_BUNDLE, thenSSL_CERT_FILE, and builds the TLS context itself. A bundle path that does not exist is an error, and verification is never silently downgraded.httpxalso inherits the macOS system proxy setting. On a dev machine that is usually a local debugging proxy (Bifrost, Charles, mitmproxy) re-signing traffic with certificates OpenSSL 3 rejects outright, so requests fail for reasons unrelated to the API. jevgrep honoursHTTPS_PROXY/ALL_PROXY(a deliberate choice) and ignores the system-wide toggle.
- Blank lines are skipped without being sent (they cost money and mean nothing).
- Results stream back in input order, with
-jrequests in flight. - Malformed JSON under
--jsonis a hard error naming the line, not a silent skip. -qstops at the first match, so it stops spending too.- The predicate is sent as a two-way
choice, not aboolean. The gateway returns no confidence for boolean answers, which would quietly make--min-confidencea no-op. - Provider overload arrives as HTTP 200 with an error body, so retries key off the message
as well as the status.
Retry-Afteris honoured on 429, capped at 30 s per record.
MIT.