We Classified 100,000 Rows in 40 Seconds: Introducing prompt_jev()

​Text classification at scale often means choosing between two imperfect options: make an LLM call for every row and watch throughput crawl, or train and maintain a specialized model. MotherDuck's new prompt_jev() function offers another path: SQL-native classification designed for fast, structured decisions, with labels and confidence scores you can immediately filter, join, and aggregate.

​In this session, Jacob Matson and Dumky de Wilde will run prompt_jev() against a live dataset, while Hamilton Ulmer explains why this workload benefits from a model built for structured decisions rather than generative text.

​We'll also unpack the benchmark behind the launch: on 100,000 rows from the AG News dataset, Jev reached 89% accuracy in 40 seconds—about 2,500 rows per second. GPT-5.6 Terra reached 88% accuracy in about 32 minutes, with a retail cost of $37.58 versus $0.50 for Jev.

​You'll see where prompt_jev() fits, how to use it from SQL, and what becomes possible when classifying an entire table no longer means waiting half an hour for the result.