Use Jev to categorize, score, and flag spreadsheet rows

Choose the column you want to add

Start with a question you need answered across your rows. You might want to group support tickets by subject, find messages requesting a refund, or see which bug reports contain enough detail to investigate. Each Column Help job adds one result column using a definition you can edit.

These questions can apply to the same spreadsheet. Run a separate job for each decision, then combine the exported columns in your working file. Keep each question specific enough that you can explain what a correct answer would look like.

Work through your first spreadsheet

Open Support ticket category and select Try example data. This loads three messages and a starting set of categories: Billing, Technical, Account, Shipping, and No match. For a larger practice table, download the 100-row retailer support CSV.

1. Preview data. Check the column names and confirm whether the first row contains headers. For your own work, paste cells from Excel or Google Sheets, including the header row, or upload a CSV. Check that each message has landed in the right cell.
2. Choose inputs. Select the fields Jev needs to read. In this example, use Subject and Message. Leave out unrelated columns and any existing answer column. A record ID may be useful in your export without being useful evidence for the decision.
3. Define results. Review the categories and name your new column. Use Advanced / Edit definition to clarify what belongs in each category. Keep No match available for a message that does not fit or provides too little relevant information.
4. Review & process. Select Preview 20 rows to try the definition on a sample. If you loaded the three example messages, all three are included. Read the source messages beside the results and inspect any rows marked Needs review.
5. Refine, then continue. If two categories overlap, edit their definitions and preview again. Changing the question requires new results. Once the definition fits your work, use Continue remaining rows if your table has more rows to process. If the preview covers your whole table, you can move on to export.
6. Review and export. Select a result to correct it, or use Accept to keep a result you have checked. Choose Copy new column, Copy full table, or Download CSV. Keep the original messages alongside the answers so you can check them later.

For your first real job, choose a small table you understand well. Include clear examples, close calls, and incomplete messages. That makes it easier to tell whether your definition asks the right question before you use it across a larger file.

What Jev does with each row

Jev is TypeSafe AI’s decision model. It takes the supplied information and a question with a defined answer format. Column Help sends the selected fields from each row with your decision definition, receives the result, and displays it with the original row for review.

Categorize

Uses Jev’s Choice question. You supply the categories and what they mean; Jev selects one. The support example turns a message about a duplicate charge into the label Billing.

Score

Uses Jev’s Score question. You define an ordered rubric. Column Help turns the returned probabilities into a score and rounds it to the nearest level on your scale. Confidence is a separate review signal.

Flag

Uses Jev’s Noul question, which returns a probability for a statement. Column Help applies your Yes and No thresholds. Results between those thresholds are marked Needs review.

Column Help handles the connection to Jev and the spreadsheet steps around it. You choose what the result should mean. These tools produce categories, scores, and flags; use a writing tool when you need a new paragraph, a translation, or a drafted reply.

The same habits help when working with other classification or decision models: supply relevant evidence, define the possible answers, and check performance on examples from your own work. Their features and uncertainty measures can differ. Column Help currently uses Jev for these decisions and does not offer a model selector.

Make the starting tool fit your work

Browse the tool catalog by type or work area. If you have already added your data, select Suggest tools in Define results. Jev uses your selected column names and a sample of up to 12 rows to suggest relevant tools. Suggestions use no row credits; select Use tool to apply one.

For categories, describe the boundary between similar labels. If Billing includes refund requests, say so. If a message mentions both a late delivery and a refund, decide which topic should determine its category. You can add a separate refund flag when both facts matter.

For scores, explain what evidence is needed at each level. The Bug report reproducibility tool measures reproduction detail. A high score says more of that detail is present; it does not establish that the bug is severe.

For flags, ask one focused question. The Refund requested tool asks whether a customer explicitly requests a refund or reimbursement. “What is your refund policy?” and “Please refund this charge” should lead to different answers. Include such close calls in your preview.

Use uncertainty to decide what to review

Jev returns answers in the requested format, but its judgments can still be wrong. A confidence value or probability is a signal for review. It is not a guarantee that a result is correct for your data. Check some apparently confident answers as well as those marked Needs review.

For Categorize and Score, the confidence threshold controls which results need review. For Flag, separate Yes and No thresholds leave room for uncertainty between them. Widening that interval sends more rows for review. Changing a review threshold uses the results you already have and does not spend more credits.

Changing a category, rubric, or condition changes the decision itself, so preview the revised definition before processing again. A threshold cannot resolve an unclear question or supply missing evidence. Rows with no usable text in the selected fields are skipped; No match is a separate Categorize result for nonempty rows that do not fit.

A refund flag helps you find a request to inspect. It does not approve or issue a refund. Use the exported column to support the review or reporting task you intended, and retain responsibility for any action you take from it.

Know what is sent and what it costs

Your full spreadsheet and results are saved in this browser. The source columns you select and your decision definition are sent to Column Help and TypeSafe AI for processing. Column Help does not persist source rows, definitions, or model outputs in its database. TypeSafe’s own data policy applies to the information it receives. Read the privacy and data handling details before choosing what to send.

Each job includes 100 successful rows at no charge, including the rows processed in your preview. Further successful rows use one credit each. Empty rows and failed processing attempts are not charged. Additional rows use prepaid credits, which do not expire; there is no subscription. You can view the current credit packs before paying.

You do not need to obtain a TypeSafe API key or set up separate Jev billing. Column Help provides access through its own row credits. Export the results you want to keep before clearing your browser’s local data or moving to another browser.