jselect
September 20, 2026 · View on GitHub
All projects · Search and retrieval
Selects source-linked evidence within a token budget using Jev Noul relevance judgments and local diversity-aware selection.
| At a glance | Details |
|---|---|
| Source | Source |
| Maintainer | keltokhy (Khaled Eltokhy); submitted from the maintainer's account with Codex assistance. |
| Format | Python library and CLI; jev-select 0.1.1. |
| Requirements | Python 3.10+; live inference requires a configured TypeSafe or OpenRouter key and incurs provider charges. |
| Access | MIT-licensed source; bring your own provider key. No hosted service is included. |
| License | MIT. |
| Disclosure | Maintainer self-submission, prepared and source-reviewed by Codex; no independent human review or directory endorsement is claimed. |
When to use
Use it to assemble evidence for a research question from logs, source files, CSV/JSONL records, or a saved local index. It returns excerpts with citations rather than a generated summary.
How it works
In semantic mode, Jev receives the task and passage text and answers one Noul relevance question per passage. Local code indexes the inputs, balances relevance against repeated text, and enforces the output token budget. Without credentials it uses a disclosed local lexical mode; that mode does not call Jev.
Inspected implementation:
Default model identifiers at the reviewed revision: jev-1.13.0 on TypeSafe; typesafe/jev-1.13 on OpenRouter.
Get started
Install uv and use the upstream instructions for other package managers.
uv tool install jev-select
# Offline example with synthetic text; no provider key or calls.
printf '%s\n' 'The signup email never arrived.' | jselect "signup problems" --local --encoding bytes --tokens 600
For live use, configure TYPESAFE_API_KEY or OPENROUTER_API_KEY privately, following the upstream README. Input text is sent to the chosen provider and calls may be billed.
After configuring a provider key, remove --local and pass --mode semantic to require Jev scoring. For example, jselect "signup problems" conversations.jsonl --mode semantic --tokens 2000. The expected output is selected source text and citation headers, not prose generated by a model.
The example commands above were checked against source and documentation. The submission review ran only the tests listed below; it did not run a live provider workflow or independently test the published package installation.
Examples and demos
- Upstream usage examples document the CLI and Python paths.
- Tests provide runnable fixture-based checks.
- No separate hosted demo was verified for this submission.
Limits and data handling
Semantic mode sends the passage text being scored and the task to TypeSafe, OpenRouter, or a configured gateway. The default semantic scan scores all eligible passages, while explicit shortlist mode can miss evidence outside its candidates. Local mode is lexical retrieval. Scores and indexes are cached locally. The output token budget does not apply to all JSON metadata. The estimated spend guard is not a provider-enforced billing cap.
Review and maintenance
Reviewed on 2026-09-20 at commit 460b217. Codex inspected the public README, license, package manifest, Jev call sites, and relevant tests.
Ran uv run --group dev pytest -q tests/test_judge_cli.py tests/test_selection.py: 59 passed in 0.93s. These offline fixture checks verify software behavior; they do not measure Jev accuracy, latency, or calibration. No live inference was performed for this listing.