What you can test
- Classifying supplied feedback text.
- Comparing topic and review timing independently.
- Testing your own rows and finite labels.
To classify text in Google Sheets with Jev, give each formula a cell, a question and allowed labels. This tutorial pairs source-compatible formulas with our separate six-row web demo.
Use one question to identify the feedback topic and another to suggest when a person should review it. The two answers describe different things: a resolved bug can remain a Product defect while receiving Review later.
Our six messages are original synthetic examples. Try the live demo at the top of this page, or use the Playground for another small classification task. The web demo makes real API requests; it does not execute Google Sheets formulas.
The formula interface comes from dbredesen's Jev for Sheets. We inspected its source at commit 6644d9544f8e8f15909f80555b18a3a8b1cd4425. We did not independently install its Apps Script project or reproduce the author's worksheet tests.

Keep the input explicit, define the choices and inspect the result.

Start with one text input, or use the six original rows in Batch classification.

Ask for Topic and Next step independently over the same feedback text.

Keep the labels explicit so your code receives a result from a known set.

Inspect both decisions and any mixed feedback before acting on the result.
The complete rules JSON defines both questions and their shared instructions. Use stated events and timing. A customer's report is a claim to investigate, and tone alone does not establish urgency.
The topic question includes Billing, Product defect, Feature request and Other. It uses Other when there is no clear main topic. The next-step question has Review now and Review later; neither label authorizes an automatic refund, reply or rejection.

The author's installation instructions describe a script attached to one spreadsheet. Execution happens on Google's servers. These are documented steps, not our installation test. The reviewed version does not have an approved public Marketplace listing.
Import the sample feedback CSV and put its text column in A2:A7. Create a Rules tab and paste the one-row Rules TSV into A1; it fills A1, B1 and C1 with the full shared instruction and two questions. The ready-to-copy formulas use those cells. Both downloads use the same rules as our demo.
For the topic column, enter =JEV_CHOICE(A2,Rules!$A$1&CHAR(10)&Rules!$B$1,"Billing","Product defect","Feature request","Other").
For the next-step column, enter =JEV_CHOICE(A2,Rules!$A$1&CHAR(10)&Rules!$C$1,"Review now","Review later"). Fill both formulas down through row 7.
These examples follow the reviewed function signature. Each question includes the full policy. The source adds instructions scoped to data and sends the cell value there. We have not executed these formulas in Sheets, so their results are not assumed to match our web run.
Load the six samples and keep both dimensions selected. They cover duplicate charges, a blocked export, a feature idea, praise, a resolved glitch and mixed feedback. Run them, read the labels beside the unchanged text, and export the results.
Then edit one message and rerun it. For example, change whether the reported problem still blocks work. This creates a new observation under the same rules; it does not establish how the model handles every customer message.
Our adapter puts one row in state.text. It builds two Choice questions with the TypeSafe JavaScript SDK. Each question's instructions contain its task, the shared classification rules and a scope statement limiting the judgment to that text.
Both questions share one systemOne call. In the author's wrapper, two formulas make separate calls using state.data. Matching business rules does not make these requests identical. Our Jev model guide explains the broader decision interface.
On September 25, 2026, all six requests in our controlled adapter run succeeded, returning 12 choices from jev-1.13.0. Download the observed results to inspect each input and response.
Median client-observed duration was 348.5 ms per request, ranging from 296 to 902 ms. These adapter timings include network and response handling. They exclude browser interaction and Turnstile, and say nothing about Apps Script execution.
The usage-derived estimate was $0.000233016 for those six requests using the recorded model rate. It excludes separate selection and review calls and is not an invoice. No expected labels were independently annotated, so we report outputs rather than accuracy.
| Sample | Topic | Next step |
|---|---|---|
| feedback-01 | Billing (100%) | Review now (100%) |
| feedback-02 | Product defect (100%) | Review now (100%) |
| feedback-03 | Feature request (100%) | Review later (100%) |
| feedback-04 | Other (100%) | Review later (100%) |
| feedback-05 | Product defect (100%) | Review later (100%) |
| feedback-06 | Billing (57%) | Review later (95%) |
Original six-row API run, jev-1.13.0, recorded 2026-09-25.Verified
The sixth message mentions a possibly high receipt and an incomplete export, without identifying a main concern. Jev selected Billing at 0.57; Other received 0.22 and Product defect 0.21. Topic confidence was 0.42. Its next-step result was Review later at 0.95.
A separate same-model review, explicitly matching each row to each question, marked 11 decisions consistent and this topic decision ambiguous. Ambiguous received 0.60 in that review. We retained the original Billing output; the review is another model judgment, not independent validation.
The Choice reference distinguishes the selected label, its probability distribution and confidence. Neither probability nor confidence measures accuracy on your customers. Read the feedback before turning a label into work.

The reviewed source adapter has no cache, cross-formula batching or automatic retry. Two formulas evaluated once on six nonblank rows imply 12 requests. Recalculation can add requests; fewer calls alone do not prove lower token use or cost.
Google limits custom functions to 30 seconds. The source stores a connected key for the spreadsheet; collaborators consume that connection's usage. Trust editors who can modify the attached script, and remove private customer details before testing.
Google also offers AI() categorization. This tutorial explores one Jev workflow without benchmarking alternatives. Keep a copy of your text and rules, inspect a small sample, then adapt the pattern using the examples library.
Understand the source workflow, live results and data handling.
It returns the selected label. The author's raw JEV function returns full response JSON, including probability fields. Our web demo also retains the distributions.
Each source formula makes its own request. Our demo sends both questions together for one row. Six rows still require six requests; uploading them does not create one combined call.
The reviewed convenience functions accept one data cell or literal. Vertical ranges can supply allowed labels, but a multi-cell data range is rejected. Fill the formulas down instead.
No. The demo policy places it in the normal human-review queue. Unclear current impact calls for clarification, not automatic rejection.
Start with six samples, inspect the returned decisions, then edit the rules for your own small test.