Convergence Loop
定義
A convergence loop is a quality control mechanism in AI workflows that links an eval step (quality gate) back to an upstream step. When the eval scores output below a configurable quality threshold, the workflow automatically re-executes the upstream steps with feedback from the eval, iterating until the output meets the quality bar or a maximum iteration count is reached.
How It Works
Add an EvalStep after any recipe or LLM step and enable convergence. Set a quality threshold (e.g., 80/100) and maximum iterations (1-10). If the eval scores below threshold, it feeds its critique back into the upstream step's next iteration as context, enabling self-correction. This creates a refinement loop that produces higher-quality outputs without human intervention.
相關術語
AI 工作流程
了解什麼是 AI 工作流程以及它們如何自動化多步驟流程。工作流程將配方與分支、迴圈、審核關卡和平行執行串連在一起。
AI Bakeoff
An AI Bakeoff is a structured comparison that evaluates multiple LLM models or prompt variations on the same inputs using automated judge scoring.
DAG Execution
DAG (Directed Acyclic Graph) execution runs workflow steps concurrently based on their dependency graph, enabling parallel processing of independent tasks.