Dspy

If you're tired of manually tuning prompts or building fragile LM pipelines, DSPy from Stanford NLP gives you a declarative framework that optimizes them automatically. You define signatures (inputs to outputs), compose modules like ChainOfThought or ReAct, then run optimizers like BootstrapFewShot to improve performance using your training data. It's designed for complex systems like RAG pipelines, multi-hop reasoning, or agents where you need reliability and maintainability. The learning curve is real, but once you get past the abstraction layer, you stop writing brittle prompt strings and start building systems that actually improve themselves. Worth it if you're building production AI workflows that need to scale beyond one-off prompts.