AdalFlow

Why we included this project

AdalFlow treats prompts and few-shot examples as code you can optimize, not strings you hand-tune. You assemble RAG, chatbot, or agent pipelines from modular components with input and output shapes declared up front, so it is always clear exactly what text reaches the model. Its trainer is the interesting part: prompts become optimizable parameters, and a single loop applies LLM-based text gradient descent along with few-shot bootstrap, improving them against your own evaluation metrics. That appeals to teams with a labeled dataset who are tired of prompt tuning as guess-and-check. The AdalComponent structure mirrors PyTorch Lightning, so anyone used to model training will find familiar ground when wiring up evaluation and loss functions.