Task guide

Should I use Jev for sentiment and tone detection?

Verdict for the sample idea

Sample idea

Detect if a customer message in our support chat is written in an angry or frustrated tone, so that we can move it to the front of the queue. It must work in real time, for about 10,000 messages a day.

Explanation

Tone detection is a clean Jev task. Each message needs one judgment. The output is a probability or a level, and it must arrive while the conversation is still open. An LLM can do it too, but it takes seconds and costs much more per message. Plain code cannot do it. A keyword list flags "this is killing me, thank you so much!" and does not flag a cold, polite cancellation threat.

Shape it as a Noul if you need one decision: "Is the customer angry or frustrated in this message?". Use the probability as the sort key for the queue. You do not need a threshold at all if you only reorder.

Shape it as a Score if you need levels. Write the levels as situations, not as degrees. TypeSafe's docs are direct on this: describe situations, because "moderately angry" gives the model nothing to match. Better levels are: neutral or positive. Disappointed but cooperative. Frustrated, repeats the problem, mentions wasted time. Hostile, insults or threats to leave. Jev evaluates each level on its own, so never write "more angry than the level before".

Send the last few messages of the conversation, not only the newest one. "Fine." means different things after different histories. This is the evidence rule, and it matters more than the wording of the question.

The common trap is to use the score as a measurement. The score is a weighted mean of level numbers, so 2.6 is not "30% more angry" than 2.0. Use it to sort and to cross a threshold, and for nothing else.

The second trap is sarcasm and mixed languages. Collect 100 real messages where your agents disagreed with the label, and test on those, not on the easy ones.

Related tasks

• Should I use Jev for support ticket routing?
• Should I use Jev for spam and content moderation?
• Should I use Jev for summaries and writing?

Verdict produced by Jev on 2026-09-19, rubric version 2. Text written by a person. How the tool works: the report.