Zet Routes Uncertain Laya Answers to Human Review

Summary

Zet is an open-source Python tool that adds a sure-or-unsure decision to Laya's text answers, routing uncertain responses for human review. It learns this routing boundary from examples with known correct answers, so future responses can either be used automatically or held for checking. The project runs locally with Python 3.10 or later and is released under the Apache 2.0 license. In a test of 350 human-labeled voice-assistant requests spanning six possible topics, Laya alone produced 45 incorrect answers. Zet routed 41 of those errors to review, while four incorrect answers still passed through. It accepted 227 answers automatically and sent 123 answers to review overall; 82 of the reviewed answers were answers that Laya had actually gotten right. No person performed the reviews in this test, so the numbers measure the routing tradeoff rather than the quality or speed of human review. The result is specific to this dataset and does not establish the same performance on other inputs, such as email. The project presents its benchmark, method, and sources alongside the local source code.