How do I know the reply is really Jev and not a random fallback?

The app separates three states and never blurs them. A key with a successful call gives a normal bubble whose hover shows the energy score and question probability; a failed call produces no answer at all, with a red error row, a 502 and counters that do not increment; no key at all gives a yellow DEMO label and a local-mode badge. A failed Jev call never falls back to the local random picker, on the server or in the browser.

Does the API key ever reach the browser?

No. The ai package and the Jev question module are imported server-side only, and the single HTML file just fetches its own reply endpoint. The README suggests checking this yourself by grepping the HTML for API_KEY, vck_ or process.env and expecting zero matches. It also warns at startup if you put a gateway-style key into the direct TypeSafe variable, which is a common mistake.

Do I have to go through the Vercel AI Gateway, or can I call TypeSafe directly?

Either works. The gateway key routes through Vercel AI Gateway, the direct TypeSafe key bypasses it, and if both are set the direct path is preferred. The model id defaults to typesafe-ai/jev and is overridable, and the punctuation rules are derived in code, so swapping the word list means editing one constant that widens both the API options and the TypeScript union.

How do I run it locally?

Clone the repo, run npm install, copy .env.example to .env and fill in AI_GATEWAY_API_KEY, then npm run dev to serve the page at http://localhost:3000.

What three questions does it ask Jev in one call?

kelime is a choice over the five phrases, enerji is a score from 0 to 2 that sets punctuation strength, and soruSayilir is a boolean probability that decides question punctuation.

How does the punctuation get chosen?

It is derived in code from the two numbers Jev returns: a high question probability gives ?, ??, ?!, or ...?, and a high energy gives !, !!, or !!!.

CommunityJev Practices & Patterns

Bes Kelime Jev

Q: What actually stops someone draining the key after I deploy this?

The exact limit lives outside the app: an AI Gateway project budget, which the README calls the real guarantee, plus a Vercel Firewall rate-limit rule on the API path. The app's own daily cap and per-IP brake are in-memory counters kept per server instance, so they are coarse brakes rather than quotas, and a cold start resets them. If you need durable counters, move the guard counter to a KV store such as Redis.

A Turkish chat toy that answers every message with one of five fixed phrases, using Jev as an evaluation model so the five-word limit is enforced by the type system rather than by prompting.

See highlights

Open github.com

Stars: 2
Language: TypeScript
License: MIT
Published by: Community
Author: mahmut-gundogdu
Last verified: Oct 2026

mahmut-gundogdu/bes-kelime-jev

User repository on GitHub

View on GitHub

Ne yazarsanız yazın, beş kelimeden biriyle cevap veren sohbet botu. Kelimeyi TypeSafe AI'ın Jev evaluation modeli seçer.

Forks: 0
Last pushed: Updated Sep 2026

Repo stats from the GitHub API, cached Oct 2026.Homepage

About this project

Ne yazarsanız yazın, cevabı hep aynı beş kelimeden biri: aynen · eyvallah · ne alaka · Sıkıntı yok · Yani. Değişen tek şey sondaki noktalama. Hangi kelimenin çıkacağına TypeSafe AI'ın Jev modeli karar verir.

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

Tagscommunity awesomejev catalog-import

Works forJev for Classification Jev for Evaluation

Used inJev for Software & DevTools

Languages: TypeScript

Highlights

Three questions are evaluated in parallel in one call: a choice for the word, a 0-2 score for energy, a boolean for questions.
Punctuation is derived in code, so the only variation is the trailing mark, from a plain period through triple exclamation.
One dependency, no build step, no framework, and the UI is a single HTML file that only posts to its own reply endpoint.
A verify script connects to real Jev and prints the outbound request, the raw JSON, token spend and repeat consistency.

Watch out

experimental_evaluate is an experimental API that may change in patch releases, and it requires ai@7. Without a key the app runs in demo mode, picking with Math.random() and labelling every bubble DEMO. The app-level daily cap and per-IP brake are in-memory counters per server instance, so the hard spend limit is the AI Gateway budget plus the firewall rule, and the README says so explicitly. Node 22 or newer is required, with 24 recommended for built-in TypeScript support. The deployed endpoint is public on the production alias; preview deploys return 401. 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-lm

Word-level language model experiment that uses Jev as its output layer, with explicit tokenizer, sampler, repetition mask, stop rules, and chunk verification.

Jev Practices & Patterns#community#jevbest#catalog-import

Community

y0usaf/jev-lmTypeScript

jevarena

Open arena pitting Jev against a chosen opponent on your question, hiding identities until you vote, then revealing latency, cost, and confidence data.