When does it chunk versus passing my document straight through?

Under about 28K estimated tokens it calls Jev directly; above that it chunks, usually at 800 tokens, scores each piece per question, refills the budget in original order with omission markers, and widens once if Jev reports insufficiency.

Which selector should I pick for scoring the pieces?

Jev reads every piece in parallel, embed_then_jev shortlists the top 25 percent with embeddings before Jev scoring, embed uses only qwen3-embedding-8b similarity, and truncate keeps the start and end.

How does provider and key selection work with three gateways?

It uses the first key found among TYPESAFE_API_KEY, OPENROUTER_API_KEY, and AI_GATEWAY_API_KEY unless LONGJEV_PROVIDER or provider names one. With TypeSafe plus embed selectors, embeddings still go through whichever of OpenRouter or Vercel has a key.

Where does my long document actually get sent and stored?

State goes to TypeSafe directly or via OpenRouter or Vercel, plus the embedding model when used. The disk cache hashes requests rather than text and the score log keeps positions and scores without text.

How do I install longjev?

uv add git+https://github.com/avshalomd/longjev, then export TYPESAFE_API_KEY, OPENROUTER_API_KEY, or AI_GATEWAY_API_KEY. It is an experiment, independent and not affiliated with TypeSafe AI.

What is the license and price?

MIT. All three providers list Jev at $0.042 per million input tokens, and an optional cache path stores every request on disk so a rerun costs nothing.

What happens if the kept text is still not enough?

If Jev rejects the request as too long, or Jev's extra sufficiency question says the kept text was not enough, longjev widens the token budget once and Jev answers again.

CommunityJev Models & Reimplementations

Longjev

Python wrapper giving Jev long-input support: scores chunks per question, keeps the best within a token budget, marks omissions, and checks sufficiency before answering.

See highlights

Open github.com

Stars: 1
Language: Python
License: MIT
Published by: Community
Author: avshalomd
Last verified: Oct 2026

avshalomd/longjev

User repository on GitHub

View on GitHub

Long inputs for TypeSafe AI's Jev decision model. An experiment.

Forks: 0
Last pushed: Updated Sep 2026

Repo stats from the GitHub API, cached Oct 2026.

About this project

Long inputs for Jev, TypeSafe AI's decision model. Jev reads at most 32K tokens. longjev has the same call as Jev's systemone(state, questions) but takes a state of any length: Jev scores every piece of the input, the best pieces are kept, and Jev answers over what is left. Status: an experiment. Independent work, not affiliated with TypeSafe AI. Jev can be reached three ways. longjev uses the first provider whose key is set, in this order, unless LONGJEVPROVIDER or LongJev(provider=...) names one:

Quoted from the project's README. Read the full README.

Tagscommunity github

Works forJev for Classification

Used inJev for Software & DevTools

Languages: Python

Highlights

Same system_one call shape as Jev with kept-piece tracking and sufficiency verdict.
Budget defaults of 16K final tokens, 28K direct-pass ceiling, and 800-token chunks.
Async asystem_one support for Jupyter plus file cache making reruns free.
Vercel gateway offers only unversioned jev, so pinned versions are not guaranteed there.

Watch out

Experiment and independent of TypeSafe; token estimates use 3 chars per token and misjudge some scripts. 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.

More Jev resources like this

gut

Python helper that turns judgment calls into one-liners like likely, classify and rate, defaulting to Jev while also running on local or OpenAI-compatible backends.