stanfordnlp/dspy

DSPy is a Python framework from Stanford NLP that changes how you build applications powered by large language models (LLMs like GPT-4, Claude, or open-source models). The central idea is to replace hand-written prompts with structured Python code, then let DSPy automatically optimize those prompts or even fine-tune model weights to maximize performance on your specific task. The problem with the conventional approach to LLM development is that prompts are fragile and labor-intensive. A carefully crafted prompt that works well on one task may break with a different model, a slightly different question, or a new version of the same model. Developers end up spending enormous effort tweaking prompt wording rather than building better systems. DSPy treats the prompt as a hyperparameter, something to be automatically tuned, rather than something to hand-craft. The way it works is through "signatures" and "modules." A signature is a short, declarative description of what an LLM call should do (for example, "given a question and context, produce an answer"). Modules are composable building blocks you chain together in Python code to build multi-step pipelines, like a retrieval step followed by a reasoning step followed by an answer generation step. Once your pipeline is written, DSPy's optimizers analyze examples of correct outputs and iteratively refine the prompts and example demonstrations that each module uses, essentially teaching the model what good outputs look like for your specific task. You would use DSPy when building complex LLM systems, question-answering pipelines, retrieval-augmented generation (RAG) systems, multi-step reasoning agents, or classifiers, where prompt quality significantly affects results and you want a principled, automated way to improve them. The tech stack is Python, installable via pip. It supports multiple LLM backends including OpenAI, Anthropic, local models, and others.