DuckDB Extension for Row Classification

Nobody measured it. The thread raises exactly that comparison and the answer is only that Jev wins when you have nothing labeled, which is a convenience argument rather than an accuracy or cost result. A fine-tuned encoder is cheaper once training is amortized.

Every judged row's content is sent to the Jev API, so the calls are billable and the data leaves your machine. That is the price of needing no labels: roughly 10 seconds per 1,000 rows is fast, but it is still 1,000 rows of your file crossing the wire.

Both. CSV, Parquet, and DuckDB tables are all named in the post, and classification drops into SQL as another expression, which is where the ergonomics argument comes from: it becomes part of the query instead of a separate modeling pipeline.

It got picked up. The author later announced that MotherDuck shipped Jev support directly in SQL as prompt_jev(), citing 50x the speed and 1% the cost of comparable frontier models in their own tests. Those are vendor numbers on their own platform, but the SQL surface is now first-party.

Embeddings treated as retrieval. Max Andrews replies that classification can go much faster that way, claiming about 20,000 rows per second on CPU without any performance work. That is roughly two orders of magnitude above the extension's own about 10 seconds per 1,000 rows, so Jev buys convenience rather than throughput.

Hamilton Ulmer, who lists himself as working on product design and engineering at MotherDuck, DuckDB's commercial cloud service. The extension is his own project rather than an official DuckDB release, though the same vendor later added native Jev support in SQL as prompt_jev().

No, it is the author's framing rather than a result. Better than using an LLM and way more ergonomic than a classifier are his words for why a decision model suits row-by-row classification, and the post publishes no accuracy, no error rate and no baseline against either an LLM or a trained classifier.