Jev Deep Research
Research agent where GPT plans and Jev finds evidence in parallel across document regions with Choice and Noul.
Overview
Jev Deep Research explores parallel evidence finding for research agents. GPT decides what to investigate; Jev evaluates many independent document regions concurrently; code returns the selected original passages for GPT to verify and use in the next step. One jev_find tool call fans out into independent Jev requests and gathers results before GPT continues.
GPT runs a BM25 search through Pi-Serini, which caches up to 1,000 ranked documents, then supplies a rank interval and up to six evidence questions. Code splits the documents into regions, and Jev answers a Choice (where is the evidence?) and a Noul (does this region address the question?) for each, in batches of 20, 40 or 60 documents that also cap concurrency. Code returns verbatim lines with document IDs within a 20,000-character budget, marking omissions and failures explicitly. Experiments target BrowseComp-Plus, and a paper is announced as coming soon.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Choice | Where in this region is the evidence for the question? | — |
| Yes / no | Does this region address the question? | — |
How it works
- GPT searches with BM25 and picks a rank interval and evidence questions
- Code reads documents and splits them into regions
- Jev judges regions in parallel with Choice and Noul
- Code returns verbatim passages; GPT verifies and continues or answers
Strengths
- Parallel evidence scanning inside one tool call
- Returns original text with line numbers, not summaries
- Explicit reporting of omissions and failed scans
Caveats
- Unofficial and experimental
- Paper not yet published
- Requires GPT access as well as Jev
Getting started
See the Quick start and Run BrowseComp-Plus sections in the README.At a glance
- Type
- Browser extension
- Status
- research
- Language
- TypeScript
- Runtime
- Pi extension (Pi-Serini)
- Has code
- yes
- License
- Apache-2.0
Use cases and integrations
How curated lists describe it
Evidence retrieval: uses Jev `Choice` to locate source lines and `Noul` to check evidence presence across document regions in parallel, then returns original passages to a GPT research agent through Pi-Serini with 20/40/60-document batch limits.
yibie/awesome-jevUnofficial experimental research agent: Jev Choice locates evidence and Noul checks its presence across document regions in parallel; Pi-Serini returns original passages for GPT to verify and continue
AnotiaWang/awesome-decision-modelsGPT directs the research; Jev uses Choice and Noul to locate evidence across document regions concurrently, and code returns original passages for GPT to verify and continue. The released integration pairs Pi-Serini BM25 search with Jev batches of 20, 40, or 60 documents.
logicrw/awesome-jev-projectsUses Jev Choice and Noul to locate evidence across document regions concurrently, returning original passages to a GPT research agent through Pi-Serini.
cobanov/awesome-jev
Frequently asked questions
What does Jev decide?
Where evidence lies in each document region and whether the region addresses the question.
What does GPT do?
It plans searches, verifies the returned passages and writes the cited answer.
What benchmark is used?
BrowseComp-Plus with Pi-Serini BM25 search.
Related projects
Web search CLI for agents that has Jev score results and scraped chunks against the query and goal, passing only relevant ones on.
BenchmarkbetaGo- Jev Search★ 534
Plain-language web search where a decision model like Jev picks query, sources and time range, then ranks Search1API results.
AppproductionTypeScript Assembles source-linked evidence for an AI task within a token budget using Jev relevance judgments and diversity-aware selection.
ProjectbetaPythonChrome extension that re-ranks Google results with Jev questions per result and folds away sales pages and SEO filler.
Browser extensionbetaTypeScriptBrowser extension that has Jev score search results for relevance, depth, SEO filler and page type, then re-ranks them with a 0-100 badge.
Browser extensionprototypeJavaScriptshadcn/ui site-search block that shows keyword hits instantly, then has Jev re-rank the top 20 by what the visitor meant.
DemobetaTypeScript