jselect
Assembles source-linked evidence for an AI task within a token budget using Jev relevance judgments and diversity-aware selection.
Overview
jselect gives an agent useful evidence within a token budget. You provide a task, such as why people give up during signup, and your data as files, directories, piped exports or Python records. It finds relevant passages with Jev yes/no relevance judgments, favours different information over repetition, and assembles a source-linked context with excerpts and citations that an agent can read directly.
No labels, predefined categories, vector database or generative model are needed. With a TypeSafe or OpenRouter key it uses semantic scoring by default; it also supports gateways and experimental local System One servers (DiffusionGemma, Laya, GLiNER2.5-Decide) that cost nothing against the spend budget. Without credentials it falls back to local lexical retrieval and says so, and a semantic mode can be required to fail instead. It is a separate package from the same author's jgrep, distributed as jev-select. It suits analysts and agent builders working with customer conversations or other text corpora.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Yes / no | Is this passage relevant to the task? | — |
How it works
- Provide a task description, data and token budget
- Shortlist candidate passages
- Jev judges relevance of each candidate
- Diversity-aware selection picks varied evidence within the budget
- Output a source-linked context with citations
Strengths
- No vector database or labels
- Diversity avoids repeated evidence
- Explicit fallback to lexical retrieval
- Local server options keep data private
Caveats
- Hosted scoring sends passages to the API
- Local servers are experimental
- Relevance is judged per passage
Getting started
uv tool install jev-select, then jselect "Why are people giving up during signup?" conversations.jsonl --tokens 8000At a glance
- Type
- Project
- Status
- beta
- Language
- Python
- Runtime
- CLI and library
- Has code
- yes
- License
- MIT
Use cases and integrations
How curated lists describe it
Research and retrieval: selects source-linked evidence within a token budget using Jev Noul relevance judgments and local diversity-aware selection.
yibie/awesome-jevSelects source-linked evidence within a token budget using Jev Noul relevance judgments and local diversity-aware selection.
logicrw/awesome-jev-projectsSelects source-linked evidence within a token budget using Jev Noul relevance judgments and local diversity-aware selection.
hellogumbo/awesome-jev
Frequently asked questions
What does Jev decide?
Whether each candidate passage is relevant to your task.
What if I have no API key?
jselect falls back to local lexical retrieval and tells you.
How is it different from jgrep?
jgrep filters lines; jselect assembles a budgeted, diverse evidence context with citations.
Related projects
Retrieval layer for Codex that finds evidence locally, optionally ranks sanitized excerpts with Jev, and returns bounded, cited evidence.
ProjectbetaPipe grep or file lists into Jev to get every candidate scored, sorted and cut at a calibrated threshold, so agents read only what matters.
ProjectbetaTypeScriptCodex search plugin that runs ripgrep and keeps only passages Jev judges relevant to the question.
PluginprototypeTypeScriptResearch agent where GPT plans and Jev finds evidence in parallel across document regions with Choice and Noul.
Browser extensionresearchTypeScriptnpm CLI that screens large agent tool outputs with Jev, returning compact typed decisions and IDs while originals stay recoverable.
CLIbetaTypeScriptDeepSeek Harness paper-reading workbench whose PDF search reranks keyword candidates with Jev, falling back to keyword order.
BenchmarkbetaTypeScript