Question 1

Fifteen recipes with strong scores, a benchmark or marketing?

Accepted Answer

Each recipe ships a runnable script, a small labelled dataset and its full live output saved under results/, so the misses are readable rather than summarized. Recipe 14's browser agent passed 5 of 6 live tasks and the failure is documented as a near-miss. The README's own framing is that these sample results do not establish production accuracy, and it publishes the model version that answered.

Question 2

Where did these accuracy numbers actually come apart?

Accepted Answer

Several places, and the recipes route on them. Support triage got the right team 34/34 but genuinely ambiguous tickets landed at 0.60-0.67 confidence and went to a person. Dedupe produced 0 wrong automatic merges only with code-checked facts, and 1 wrong merge without them. Paging on-call on a single severity score caught 4 of 7 critical events; splitting it into four yes/no questions caught 7 of 7.

Question 3

Will it hand back an answer that is not one of my options?

Accepted Answer

It never writes text, so there is no path to an invented value outside your option list; answers are typed choice, noul and score. It can still pick the wrong option, which is why every recipe routes on confidence and measures its errors. The failure mode to watch is literal reading: on a real bot catalog a plain topic question tagged 36 of 100 templates Generative AI, and one sentence saying that does not count brought it to 6.

Question 4

Why OpenRouter instead of the TypeSafe API, and what will bite me?

Accepted Answer

The recipes call OpenRouter's /api/alpha/decisions endpoint; the chat endpoint rejects Jev. The model id ~typesafe/jev-latest needs the leading tilde, which the README flags as the second thing that trips people up. Recipes 01 to 13 and 15 need only Node 22.9+ and an OpenRouter key with no dependencies; recipe 14 adds Playwright Chromium and recipe 15 needs your own Google OAuth client.

Question 5

Which recipes need setup beyond Node and an OpenRouter key?

Accepted Answer

Recipe 14, the browser agent, needs npm install and a Chromium via npx playwright install chromium, or a CHROME_PATH. Recipe 15 needs your own Google OAuth client, though its sample inbox works without one.

Question 6

What endpoint and model id does the cookbook use?

Accepted Answer

The endpoint is /api/alpha/decisions, not the chat endpoint, and the model id is ~typesafe/jev-latest including the tilde, which always points to the newest Jev.

Question 7

How much do the cookbook recipes cost to run?

Accepted Answer

About 2 to 8 cents per 1,000 items. Jev is $0.042 per million input tokens with output free, and npm run all runs every recipe except the browser agent for about 2 cents.

Jev Cookbook — nexibeo

About this project

Highlights

Quickstart

Questions about this entry

From the Jev guides

From the Jev AI model community