วิธีใช้ Jev ใน Agent จริง 10 ขั้น พร้อมตัวเลขที่ตรวจสอบได้

The article discusses how to effectively use Jev in real-world agents, providing 10 steps and concrete data to verify. Jev, developed by TypeSafe AI, is designed to ask specific type questions rather than generate text or answer in conversational form. Unlike traditional LLMs that write text and answer in dialogue, Jev returns typed results with associated probabilities for each choice.

The article outlines a three-part process for using Jev in an agent system: separating tasks from decision-making, defining primitive types with questions, choices, probabilities, and confidence scores, and maintaining a state that includes objectives, rules, background information, and completed tasks. The key insight is that Jev makes decisions based on the current state and then executes code, with state changing after each action.

Empirical data shows that combining Jev queries into a single query significantly outperforms querying each question individually in terms of speed and accuracy. For example, when asked 13 questions about GDPR using Wikipedia, submitting all questions at once was 12.2 times faster and produced equivalent results. The article provides specific data points, including response times, accuracy, and re-ranking improvements in legal top-10 rankings.

However, it also notes that these numbers come from demonstrations rather than independent third-party verification.