stanfordnlp/dspy

DSPy solves the problem of fragile and labor-intensive prompt engineering by introducing a declarative programming model where prompts are treated as trainable parameters. It automates the optimization of prompts and weights, making AI systems more robust and adaptable across tasks and models, unlike traditional approaches that require manual tweaking for each use case.

Developers can build a wide range of AI systems, from simple classifiers to sophisticated RAG pipelines and agent loops, with automatic optimization of prompts and weights. Real-world use cases include knowledge-intensive NLP tasks, multi-stage language model programs, and self-refining pipelines with computational constraints. The ceiling is high, as DSPy supports fine-tuning and prompt optimization together, enabling state-of-the-art performance on complex tasks with minimal manual intervention.