Batch text classification without losing failed rows

Import and preview each row

Import a UTF-8 CSV, TXT or JSONL file up to 2 MiB and 1,000 data rows, or paste one message per line. Select the input column for structured files and inspect the preview before running. Each message must contain 3–6,000 characters; rejected rows are reported and never sent.

Every row in a run is judged by the same rule set: one label list, or one to four typed questions. In labels mode, include an other or unclear outcome when inputs may fall outside your categories. A choice answer still selects one of the labels you wrote, and the threshold is application post-processing — not model abstention, and not a guarantee of accuracy.

Open the batch workspace · Request limits and credit rules

One label list, or up to four typed questions

Questions mode asks one to four typed questions about every row: a choice over 2–12 criteria, a noul over true and false criteria, or a score over 2–7 ordered levels. Each answer is typed in the export, and the threshold is applied per answer — to the probability of a noul, to the confidence of a choice or a score — so needsReview in a row means at least one question was uncertain.

A preset fills the editor with a ready rule set, and the AI assistant writes a draft from a short prompt: the model proposes a name and typed questions that you review and edit before any row is sent. Generating the draft costs one credit and is refunded when it fails; each row still costs its own credit, and nothing runs until you press Run.

LLM as a judge: typed rules you can audit · Batch workspace

Know what consumes credits

Each successful row uses one decision credit whether the run asks one label question or four typed ones, and a rule set generated by the assistant costs one credit of its own. Failed account requests receive a matching refund entry. The browser submits rows sequentially, so you can stop between rows and avoid sending the rest of a batch after an account or quota error.

Export safely

Export all rows, including failures, rejected inputs and unconfirmed requests. Formula-like cells are escaped to reduce spreadsheet formula injection risk. The CSV contains row numbers, decisions, confidence, review state and errors, but no submitted source text. Keep the original file in your own system to match results back to records.

Parsing happens in your browser. Running sends the selected text and rules to Jev API Pro and its model provider; local parsing does not mean local inference. Keep this tab open and export before leaving. This is an interactive batch tool, not a background job queue or an unlimited bulk endpoint.