AdalFlow

AdalFlow, from the SylphAI team, takes the position that building LLM apps should feel like building neural networks in PyTorch. You compose Generators, Retrievers, and other components into pipelines, declare prompt templates as Parameters, and then let the Trainer run textual-gradient optimization over them — the same family of ideas as TextGrad and DSPy, unified in one framework with few-shot bootstrap optimization in the same loop.

That makes it a good fit for people who want to tune prompts and workflows systematically rather than by vibes: define the pipeline, attach a loss-like objective, and let the optimizer search. It covers chatbots, RAG, and agents, ships with benchmarks and tutorials, and is MIT licensed. With around four thousand stars it is smaller than the giants in this batch, but it is actively maintained and the PyTorch-style API is genuinely pleasant if that mental model is yours.

1. Run pip install adalflow.
2. Set your model provider API key.
3. Compose components (Generator, Retriever) into a pipeline with Parameter prompts.
4. Optionally run the Trainer with a textual-gradient optimizer to auto-tune prompts.