User repository on GitHub
GPT drives research. Jev finds evidence in parallel.
- Forks
- 1
- Last pushed
- Updated Sep 2026
Repo stats from the GitHub API, cached Oct 2026.
CommunityJev Agents & MCP
JevDeepResearch
A research agent where GPT picks what to look for and Jev judges many document regions in one call, so the code can hand back verbatim passages for verification.
- Stars
- 4
- Language
- TypeScript
- License
- APACHE-2.0
- Published by
- Community
- Author
- sunyasheng
- Last verified
- Oct 2026
About this project
GPT drives the research. Jev finds evidence in parallel. Paper on arXiv — coming soon. Quick start · Method · Results · Run BrowseComp-Plus · Docs. Jev Deep Research explores parallel evidence finding for research agents. GPT chooses what to investigate. Jev evaluates independent document regions concurrently, and code returns selected original passages for GPT to verify and use in the next research step. Our BrowseComp-Plus experiments pair Pi-Serini's BM25 search with Jev batches of 20, 40, or 60 documents.
Quoted from the project's README. Read the full README.
- Languages
- TypeScript
Highlights
- One jev_find call fans out into independent Jev requests that each combine a Choice question for location with Noul questions for relevance.
- Pi-Serini BM25 does the retrieval; GPT supplies a cached search id, a rank interval and up to six independent evidence questions.
- Returned evidence carries document ids, line numbers and verbatim text inside a 20,000-character budget, with omissions and failed scans stated.
- The Python package ships a QA reference implementation whose offline demo runs on three invented documents with no API keys or model downloads.
Watch out
The reported accuracy numbers are small development experiments on 20 questions with one attempt per question per variant, and the README says candidate width and Jev concurrency were changed together with no same-depth no-Jev batch control. That means the gains cannot be attributed to Jev rather than to the wider workflow changes. Costs use frozen token rates, exclude local compute and are not account bills, and missing usage is not imputed. The offline demo's answer and confidence come from a synthetic fixture and are not measurements of anything. The best row is a frozen experiment entry, and the arXiv paper is described as still forthcoming, so CITATION.cff is the current citation path. Running the benchmark itself needs Azure OpenAI with Entra auth, a Typesafe key and a corpus search service behind a Qwen3-Embedding-8B dense index. Copy reviewed against the repository README on 2026-10-04, not from running it.
Questions about this entry
Short answers drawn from the linked project or page.
You cannot tell from these results, and the author says so. Candidate width and Jev concurrency were varied together and there is no same-depth no-Jev batch control, so the improvement is attributable to the parallel-evidence workflow rather than to Jev specifically. Read the table as a direction, not an ablation.
The offline demo, yes: pip install -e . then jev-research demo exercises the released controller with three invented documents and scripted model responses, needing no API keys or model downloads. The BrowseComp-Plus runner is different, requiring Azure OpenAI with Entra authentication, a Typesafe key for Jev and a corpus search service with its own dense index.
Retrieval stays with BM25 in Pi-Serini, and GPT picks the search id and rank interval. Code then reads those documents sequentially and splits them into regions; Jev's job is to judge regions concurrently for every supplied question. The parallelism happens inside one evidence-finding call, while dependent research steps remain sequential.
Omissions and failed scans are reported explicitly rather than silently dropped, and merged line ranges are read back inside a 20,000-character source budget. For lines too long for the frozen reader, an optional splitter lets Jev inspect them while returned evidence keeps the original line numbers and source offsets.
The number sets both the document cap per batch and the concurrent Jev request cap. The small 20-question experiment tested all three, and Jev-60 had the best correct count, but candidate width and Jev concurrency were changed together, so the variants are not a clean isolation of either.
It ships a separate QA reference implementation: import ResearchAgent and SearcherClient, then agent.run(question_id, question) returns a result whose answer_text you can print. Direct Python use does not load .env automatically, so export the model environment variables first, and you can plug in your own corpus through the searcher interface.
Each run writes a configuration and source manifest, per-question records, a decision trace and a model-call ledger, and --dry-run inspects a command without contacting services. The code is Apache-2.0, while external datasets, models and services keep their own terms.
More Jev resources like this
48GitHub stars
Agent harness for work that outlives one chat, with resumable turns, routed state machines, and observational memory across sessions and processes.
Jev Agents & MCP#community#jevbest#catalog-import
Community
dzhng/duet-agentTypeScript
dzhng/duet-agentTypeScript37GitHub stars
Assistant framework that stores specs, memory, and specialist agents as plain markdown files in your repo for Claude Code, Codex, Opencode, and similar runtimes.
Jev Agents & MCP#community#jevbest#catalog-import
36GitHub stars
Pi extension that strips the full Agent Skills catalog from every request and loads only the skills a Jev ranking call judges relevant.
Jev Agents & MCP#community#jevbest#catalog-import
102GitHub stars
Behavioral code search that asks Jev to score repository fragments for a plain-language question and returns exact excerpts with paths and lines.
Jev Agents & MCP#community#jevbest#catalog-import
Community
nassim-arifette/jevgrepTypeScript
nassim-arifette/jevgrepTypeScript18GitHub stars
Local engine with TypeScript and Python SDKs that runs Claude Code and Codex sessions as a warm team with a durable mailbox.
Jev Agents & MCP#community#jevbest#catalog-import
Community
masonlee39/Multi-AgentTypeScript
masonlee39/Multi-AgentTypeScript16GitHub stars
Single-command harness that turns a plain-language stack description into synced docs plus blocking hooks for Claude Code, Codex, and Pi, with deterministic presets or one Jev chooser call.
Jev Agents & MCP#community#jevbest#catalog-import
Community
kyu1204/oh-my-harnessTypeScript
kyu1204/oh-my-harnessTypeScriptFrom the Jev community
Posts from builders shipping with Jev right now.