DSPy: programming, not prompting

The problem DSPy tackles

When you build an app on a large language model, you can lose hours tweaking fragile prompts: change one sentence or upgrade the model, and quality drops. DSPy, which started at Stanford NLP in late 2022 and is now driven by a large open source community, offers another route: write modular Python code and let algorithms work out the best way to talk to the model. Think of it as automated prompt engineering.

Three ideas to grasp

• Signatures: you declare the inputs and outputs of a step (say, a support ticket in, an urgency level and a team out) without writing the prompt.
• Modules: interchangeable strategies for running a signature, such as Predict (direct answer), ChainOfThought (step-by-step chain-of-thought reasoning) or ReAct (reasoning plus tools).
• Optimizers: given examples and a metric, they automatically tune the instructions and examples sent to the model (MIPROv2, BootstrapFewShot, GEPA and others), and can even fine-tune model weights.

Typical uses

DSPy works for anything from a simple classifier to a multi-step RAG pipeline or an agent loop. Common goals: moving to a cheaper model without losing quality, making information extraction more reliable, or keeping a system reproducible when you switch providers. Model calls go through a layer that supports many providers, which makes those switches easier.

Getting started

Install the dspy package with pip (Python 3.10 or later), configure the model you want to use, then write a first signature and run it through a module. The next step is to gather a small set of examples and a metric so you can run an optimizer. The official tutorials follow exactly that path.

Strengths and limits

Strengths: a rigorous, measurable approach, model independence and a foundation in published research.

Limits: the learning curve is real, since you have to think in signatures and metrics; optimization burns through many model calls; and DSPy provides no UI, no hosting and no ready-made multi-agent orchestration.

Alternatives

LangChain and LangGraph for orchestrating agents, Pydantic AI for typed outputs, and Haystack for production RAG pipelines.