DSPy

DSPy is an open-source Python framework for building and optimizing language-model systems. Developers define typed inputs and outputs through signatures, then compose modules for extraction, reasoning, tool-using agents, and multimodal tasks. Its optimizers use examples and a scoring function to tune prompts, making task definitions reusable across different execution strategies. The approach treats AI applications as modular programs whose behavior can be evaluated and improved.

Created by Omar Khattab, now an MIT assistant professor, DSPy originated at Stanford NLP, with an open development lineage dating to 2022. Its research contribution is a programming model that separates declarative language-model calls from their implementation and compiles pipelines against a chosen metric. The integrated GEPA optimizer extends this approach through collaborative research: it reflects on execution traces and textual feedback to propose prompt changes, then combines complementary candidates through Pareto-based search.

In 2026, the project reported 6.6 million-plus monthly downloads and more than 452 contributors. By August 2026, cumulative downloads exceeded 55 million. Production applications include Shopify’s metadata extraction, Dropbox’s Dash relevance judging for ranking and evaluation, and Databricks’ language-model judges, retrieval-augmented generation, and classification.