Question 1

Is this a library I can depend on, or a demo I have to copy from?

Accepted Answer

It is a reference implementation to copy structure from. The README is explicit that the playground is a demo instrument: no database, no history beyond sessionStorage statistics, and each use case is a self-contained module under src/lib/jev/usecases declaring its own state builder, questions, normalization, thresholds and examples. The reusable idea is the shape — one SDK boundary, deterministic rules in application code, model output only as typed probabilities — and docs/architecture.md has the detail.

Question 2

Does Jev get to decide the routing, or only advise it?

Accepted Answer

Only advise. The division the project states is that Jev judges and the application decides: Jev returns an intent category, a recommended handler, an urgency score and escalation or ambiguity nouls, while the code chooses which queue receives the request, when to flag review and when to ask for clarification. Thresholds such as flagging review at an escalation probability of 0.5 are application values, never model settings, and compare-and-choose ranks by a weighted composite the application computes — deliberately not a model-produced best-product score.

Question 3

How do I see what was actually sent to the API?

Accepted Answer

Toggle developer view and each evaluation shows the exact state and questions sent, the raw assessment, the normalized typed judgment alongside the application thresholds that fired, plus model, request ID, TypeSafe latency, total latency, token usage and error classification on failures. Every block has a copy-to-clipboard control, and secrets never appear in that panel.

Question 4

Can I run the tests without an API key, and does it need Vercel?

Accepted Answer

The tests are deterministic and need no credentials: node:test over the core plus each use case's normalization and rules, covering the parse layer, thresholds and malformed or partial responses through a mocked answer boundary. For running the app itself, npm install plus a .env.local with TYPESAFE_API_KEY and npm run dev serves it on port 8080; the Vercel path is documented separately as import, add the environment variable, redeploy. One bad input never poisons a triage batch, since each item carries its own success or failure.

Question 5

What limits are documented, and does it take images?

Accepted Answer

Four, plainly stated. Jev evaluates text states only, with no images, audio, or video. English is the primary training language, so other languages score lower. Probabilities are calibrated across prediction groups rather than guaranteed for any single answer. And this is a demo instrument: no database, no history beyond session statistics in sessionStorage.

Question 6

Where does the batch surface live and what are its limits?

Accepted Answer

In Content Triage, toggling Batch takes one message per line with a limit of 20, and the server runs the same pipeline per item in parallel. Each item carries its own success or failure, so one bad input never poisons the batch, and results sort by priority and filter to review-needed items. Compare & Choose instead fans out one request per candidate.

Question 7

Is there leftover cruft I should know about before forking?

Accepted Answer

Yes, and the README flags it. The environment that produced the branch could not delete superseded files, so a few forwarder shims remain under src/lib/jev/ and src/components/, marked do not import. They are inert and safe to delete, but they are there rather than cleaned up.

Jev — haibt163

About this project

Highlights

Questions about this entry

More Jev resources like this

From the Jev guides

From the Jev community