Question 1

What is the jev_naics project?

Accepted Answer

A Python library and FastAPI service from jmurray10 that classifies a business's description of operations into one of 1,012 NAICS 2022 industries, or holds it for a person, built for insurance applications where the code routes the submission to an appetite and rating class.

Question 2

Why is there no LLM in the jev_naics runtime path?

Accepted Answer

The agent is specified as PEAS and code does almost all the work: Postgres full-text search over the official Census text, spelling correction, a modern-terms lexicon, a tree walk, and every gate and hold. Jev only answers yes/no checks and closed choices, so it cannot invent a code it was never offered.

Question 3

Which classification methods does jev_naics offer?

Accepted Answer

Three, over one set of gates: hybrid (recommended) searches first and adds tree-walk finalists when search holds, using about 1 Jev call (about 7 on a fallback); search picks among the top 10 with 1 call; tree walks the NAICS tree from the 20 sectors with about 5 calls.

Question 4

What happens without a TypeSafe API key?

Accepted Answer

The code replays answers recorded on the real model from backend/transcripts, so the test suite, every eval and the page's examples still run; a description that was never recorded is refused rather than guessed, with the library holding it as service_error and the API answering 409.

Question 5

How did the hybrid method score on its eval sets?

Accepted Answer

On the blind held-out set the hybrid scored 15 right and 1 held across 16 descriptions, against 8 right for search and 14 for the tree walk; it took 32 right and 1 correct hold on dev (33), 18 right on messy dev, and zero wrong codes across all sets.

Question 6

What did the external ExioNAICS check find?

Accepted Answer

60 random unique descriptions from the public ExioNAICS dataset, run live once and judged blind: 40 got a code (36 defensible, 4 arguable, 0 indefensible) and 20 were held (15 reasonable, 5 should have acted). The dataset's own labels were wrong for 18 of the 60, so label match undercounts.

Question 7

When does jev_naics hold a description for a person?

Accepted Answer

When the description is too vague, when the pick is under the probability bar, or when the picked industry fails its fit check; the gates check the picked industry's probability and fit, never the model's own confidence, and a missing, NaN or out-of-range answer also holds.

Jev Naics

Highlights

Questions about this entry

Reactions & coverage

From the Jev guides

From the Jev AI model community