COOLJEV / PRACTICAL GUIDE
Batch review classification with Jev: labels and human review
“The mug keeps drinks warm, but the lid leaks and the delivery box arrived crushed.” This review contains two separate problems. A single “negative” category hides the distinction between product quality and shipping. The CoolJev workspace proposes labels, then lets you inspect the original, correct the labels and export the results.
Define the decision you need
This workflow helps organize product feedback and prepare issue lists for the responsible teams. It does not establish whether a review is authentic, issue refunds, remove criticism or judge the person who wrote it.
| Question | Matching evidence | Different evidence |
|---|---|---|
| Product quality | Leaks, failure, peeling coating | A disliked color alone |
| Delivery and packaging | Late arrival, crushed packaging, missing delivery | Failure after a week of use |
| Usability | Difficult cleaning, opening or fit | A question without a usage experience |
| Enough information | Concrete experience or an explicit report of no issues | “Nice,” “looking,” or “how do I buy?” |
These are independent decisions. One review can have quality, delivery and usability labels together. “No reported issue” and “insufficient information” represent different situations and should not be collapsed into one category.
From text to a suggestion
These examples are authored explanations, not live model outputs:
Input: The lid leaks and the shipping box was torn.
Expected labels: Product quality; Delivery & packaging
Next: Inspect the original, confirm the labels and prepare separate issue lists.
Input: Used it for two weeks with no issues.
Expected label: No reported issue
Input: Looks nice. I have not used it yet.
Expected label: Insufficient information
The workspace sends the material as state and asks several independent noul questions. The API returns values from 0 to 1, not a measured accuracy percentage. Noul does not have a separate confidence field. See the TypeSafe API reference.
CoolJev currently proposes an issue label at 0.8 or above and flags values between 0.2 and 0.8 for closer review. The information-sufficiency signal must also reach 0.8. These are initial workflow thresholds, not evidence of accuracy on your products, language or customers.
Process a batch
- Open the review workspace. Paste one review per line or import a UTF-8 CSV containing a
reviewortextcolumn. - Inspect the text, confirm that you may send it to TypeSafe, and select “Classify unlabelled items.” Opening a page, importing data and editing labels do not send model requests.
- Items run sequentially. Each item costs one credit. Its independent questions share one request; multiple labels do not cost extra credits.
- Inspect suggestions, particularly insufficient information and uncertain signals. Suggested results remain pending until you confirm them.
- Download CSV for spreadsheet work. Download project JSON to preserve originals, model suggestions, request IDs and review progress for later restoration.
Guests receive 10 daily credits; signed-in users receive 20 total. Credits reset at midnight China time (UTC+8), do not roll over and do not stack with guest usage. A shared network limit helps prevent abuse. Credits cannot be purchased during this beta. For larger sets, start with a small sample or run the downloadable Python examples in your own environment.
Failures, stopping and reruns
Stopping prevents later items from being sent; a request already sent may still complete. Retry failed items without repeating successful ones. When a network outcome is uncertain, the workspace reuses the original request ID to retrieve its result and avoid another debit. A confirmed failure can be submitted as a new request.
Human review and exports are free. Importing the same material again and clearing its labels starts a new classification task and consumes credits. Save a project before refreshing: page memory is not a long-term backup.
Check the criteria on your own data
Manually label a small set of representative reviews that you may submit. Include overlapping issues, short statements without experience, praise, negation and domain-specific terminology. Examine missed issues separately from incorrect labels, then improve question definitions or review procedures. Do not tune and evaluate on the same examples.
TypeSafe currently describes English as Jev’s strongest language; test other languages on your own workload. Model and language documentation
Read question design, confidence and evaluation, or try the free review exercises. For customer requests rather than product reviews, use ticket triage.
Independently written by CoolJev. Sources checked and updated: 2026-09-21