DSPy

Overview

DSPy from Stanford NLP flips the prompt-engineering paradigm. You declare modules with input and output signatures, an objective, and a dataset. DSPy compiles prompts (and optionally weights) against the objective. The result is reliable, optimisable LLM pipelines that don't depend on someone's clever wording.

Best for

Best for
Teams who want to optimise their LLM pipelines like code, not edit prompts forever

Use cases

• Build a multi-stage pipeline that can be re-optimised when the model changes
• Replace fragile prompt chains with declarative modules
• Run prompt-and-weight optimisation against a real eval
• Standardise LLM workflows across a research team

Notes

Why it matters

Prompt engineering does not compose and does not optimise. DSPy is the clearest answer to “what does programming an LLM pipeline look like when you treat it like software.”

How teams use it in production

Start with a small, evaluable pipeline. Build the eval first. Then run DSPy optimisation against it. The eval is the asset; the pipeline is the artefact.

What to watch

DSPy’s ideas are leaking into every other framework. Whether the framework itself wins, or the patterns get absorbed, the declarative shift is real.