rowlab

Ask a question of every row.

rowlab labels, extracts and rewrites tabular data with AI. Add a CSV, JSON or Parquet file, write the question, try it on twenty rows, then run it on a million. Open source, runs on your laptop, or hosted when you'd rather not run it.

Try it hostedRun it yourself
uvx rowlab

One process, one SQLite file, one data folder. Open localhost:8000, add a file, and bring your own OpenRouter key.

Try before you pay

See the rows, calls and cost before a run. Run two models on the same rows and compare them, against your own labels if you have any. Cheap decision models for yes/no questions, LLMs for the rest.

Nothing changes in place

Every run makes a new dataset with the answers added, named after where it came from. Keep the tree, export any branch as CSV, JSONL or Parquet, to a file or S3.

Built for people and agents alike

A fast web app, a plain API, and an MCP server with skills that teach agents to label well. The same work, whichever way you do it.

What a run looks like

Pick the rows with SQL, write the question from the row's own columns, choose the answer type, and a model.

SELECT * FROM data WHERE stars <= 2 LIMIT 20

Is this review about delivery?
Review of a kitchen product: {{review}}

yes / no  ·  typesafe/jev-1.14  ·  20 calls  ·  $0.00

Read the answers next to the rows. Fix the prompt. Then take the LIMIT off.

Who it's for

  • Analysts with a column to fill in.
  • Teams making training data.
  • Researchers coding open-ended answers.
  • Ops triaging tickets and reviews.

Answer types

Classify

Yes/no, pick one, pick any, or a score on a scale. Any model, each answer with its probability.

Extract

Typed fields pulled out of the row: text, numbers, dates, lists. Empty when the text doesn't say.

Write

New text from the row: a summary, a translation, a rewrite.