Question 1

Is this measuring real feelings or just pattern-matching text?

Accepted Answer

Pattern-matching text, and the README says so directly: the numbers are model judgements, not proof of what the other person feels. What you get is structured rather than generative — each line from them gets an emotion and an intent, your replies get a rating, and the exchange collapses into a single score plus a suggested next step. It deliberately does not write a 'you two would be great together' essay. The scenario selector is where real judgement sits: scoring for romantic interest, for colleagues and for social platforms uses different criteria.

Question 2

What happens to the chat text I paste in?

Accepted Answer

It is sent to whichever model provider you configured — TypeSafe, Vercel AI Gateway or OpenRouter — and billed to that key, which is read server-side only. Cached analyses do not re-hit the model, and switching which side you are and switching back reuses the last local result rather than paying twice. The part worth thinking about first is the admin view: records synced to the server are visible to an administrator, so the README's advice not to paste highly sensitive content is not boilerplate.

Question 3

Does it handle pasted exports from other apps, or only WeChat?

Accepted Answer

It handles WeChat desktop multi-select copy, QQ and WhatsApp exports, and WeChat timestamps that include the afternoon marker. The parser is defensive about a common failure: a date line is not treated as a third participant. When it cannot identify exactly two people it refuses to guess — the start button stays greyed out and the page states the reason, which is the behaviour you want from something that will otherwise assign a confident score to the wrong conversation.

Question 4

What are the redeem codes and the admin panel for?

Accepted Answer

The admin panel is the operator side of the app. You can browse already-synced conversations as bubbles with the same labels the front end shows, trigger analysis for a record that has none, and export a conversation as plain text or a single long screenshot. Redeem codes are managed there too, and only take effect once the redemption flag is enabled — with it off, analysis runs are not counted. The admin username and password ship as defaults, so change them before exposing a deployment to anyone.

Question 5

What are the install and run steps, and which ports does it use?

Accepted Answer

Node 22.12 or newer. Clone the repo, run npm ci, copy .env.example to .env and fill in JEV_API_KEY, then npm run build and npm start, and open 127.0.0.1:3178. For UI work npm run dev serves the page on 5178 while the API stays on 3178. A sync should not overwrite the server's .env or data directory.

Question 6

How does the scenario selector change how it scores?

Accepted Answer

The rubric follows the relationship scene you pick: romance, friends and colleagues, or social platforms, so a mentor's concern is not read as romantic interest. Their bubbles get emotion and intent, yours get a reply rating, and the whole exchange collapses into one favourability score with a suggested next step.

Question 7

What are the hard limits on what it can read?

Accepted Answer

Only two people's text is processed. Images and voice show as placeholders and contribute nothing to the score, and the model has no access to tone, expression, or time spent together, which is most of what actually carries the signal. The README is explicit that the output is a model's judgement about a person, not a reading of their feelings.

Jev Crush

About this project

Highlights

Questions about this entry

More Jev resources like this

From the Jev guides

From the Jev community