DSPy

DSPy, short for Declarative Self-improving Python, is a Stanford NLP framework for programming language models instead of prompting them by hand. Developers express an AI system as compositional Python modules whose behavior is declared through signatures, typed specifications of inputs and outputs, rather than brittle prompt strings. DSPy optimizers then compile the pipeline, automatically refining the prompts or even fine-tuning model weights against a metric and a small set of examples, so quality improves systematically rather than through manual trial and error. The approach fits simple classifiers, retrieval-augmented generation pipelines, and agent loops alike, and the same program can be re-compiled when switching to a different model. The project was created by Omar Khattab with collaborators including Matei Zaharia and Christopher Potts, is installable from PyPI, and is open source under the MIT license. Its users are researchers and engineers who want reproducible, optimizable LLM pipelines.