Is word-by-word Choice a sensible way to build a chatbot?

No. The digest calls it a toy inversion: Jev is built to answer batched questions in parallel, while Grug burns a grouped tournament per sequential word. Use it to learn the Choice primitive, not as a chat architecture.

Do I need an API key just to load the demo?

No. Without TYPESAFE_API_KEY it runs a deterministic mock so contributors can inspect the interface immediately. Set the key only when you want live probabilities and confidence in the inspector.

Why cap the vocabulary and stop replies so quickly?

The vocabulary caps the decision space so each tournament stays small, and replies stop at the end option or after 18 words. Recent and repeated words are withheld to break repetition loops.

How do I run it without leaking my API key?

Put the key in server-side .env, never in a public variable, and let the browser call the project server route. Only that route talks to TypeSafe, and the inspector shows accepted and rejected candidate words.

How does a word get chosen from the 979-word vocabulary?

Every eligible word gets a chance through parallel groups of at most 249 options, then Jev picks among each group's strongest finalists or the dedicated end option. The conversation and words-so-far are carried as structured state, and replies stop at end or after 18 words.

How do words from my message join the vocabulary?

The app extracts candidate names, numbers, and terms from the user message, then batches one Noul judgment per non-numeric candidate to decide which words join the conversation vocabulary. Numeric tokens are always admitted, and the inspector lists accepted and rejected words.

What happens when the API is rate-limited or returns something odd?

Every returned choice is validated, and documented 429 and 529 failures are retried. Without a key the app falls back to a deterministic mock whose answers depend only on the prompt and conversation length, which keeps the demo inspectable and testable.

CommunityJev Tools & CLIs

Jev Grug

Playful chat demo that makes Jev hold a conversation one word at a time by picking from a compact 979-word vocabulary through repeated Choice calls.

See highlights

Open github.com

Stars: 4
Language: TypeScript
License: MIT
Published by: Community
Author: mkotlikov
Last verified: Oct 2026

mkotlikov/jev-grug

User repository on GitHub

View on GitHub

Helping JEV speak <3

Forks: 1
Last pushed: Updated Sep 2026

Repo stats from the GitHub API, cached Oct 2026.

About this project

TypeSafe says Jev isn't an LLM. Let's talk to it anyway. Grug is a tiny cave-chat experiment that makes Jev hold a conversation by repeatedly choosing its next word from a compact vocabulary. What could grug wrong? The demo now connects to the real Jev API when TYPESAFEAPIKEY is configured. Without a key it falls back to a deterministic mock, so contributors can still run the interface immediately.

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

Tagscommunity jevbest catalog-import awesomejev

Works forJev for Classification

Used inJev for Software & DevTools

Languages: TypeScript

Highlights

Groups of at most 249 options tournament down to finalists or a dedicated end option, stopping after 18 words.
Carries conversation plus words-so-far as structured state and admits numeric tokens automatically.
Extracts candidate names and terms from user text and judges each with a batched Noul before admission.
Shows live probabilities and confidence in an inspector, with a deterministic mock when no key is set.

Watch out

This is a toy inversion of Jev's intended parallel use, stated as not a recommended production chatbot architecture. 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

jev-recall

Python memory filter that judges every stored memory against the current request with one calibrated yes/no per memory in a single Jev request, keeping everything above a threshold.