Trace Judge

#9 v Observabilita a evaluace

Langchain · 5× · naposledy 01. 7. 2026

44

Momentum

Trace Judge is a fine-tuned evaluation model (based on Qwen) developed jointly by LangChain and Fireworks AI that automatically detects "Perceived Error" (errors in agent trajectories) in production traces. According to LangChain, it achieves the accuracy of Frontier models (Haiku, Sonnet, Opus, GPT-5.5) while running up to 100 times more cost-effectively. The model is still in early testing (sign-up for select customers), with a broader rollout announced for the coming weeks/months. A dedicated product website, pricing list, or public API documentation does not yet exist—the only official source is the LangChain blog post/research announcement.

Zdroje (5)