DSPy

A framework for programming (not prompting) LLMs — declare signatures and modules, then let an optimizer compile prompts and few-shot examples for you.

At a glance

Field	Value
Category	LLM programming framework
Difficulty	Advanced
When to use	Tasks with examples and metrics to optimize
When not to use	Ad hoc prompting with no labeled data
Alternatives	LangChain, TextGrad, hand-tuned prompts

What it is

DSPy (from Stanford NLP) treats LLM pipelines like programs. You declare Signature types (input fields → output fields) and compose Modules (Predict, ChainOfThought, ReAct). A compiler then searches — using your training examples and metric — for the best few-shot demonstrations, prompt templates, or fine-tuned weights. The net effect is “stop hand-crafting prompts; let the optimizer handle it”.

When we reach for it at Ephizen

Multi-step pipelines where individual prompts are tangled and brittle.
Tasks where we have a good eval set and want reproducible improvements over time.
Porting a prompt-heavy pipeline to a smaller, cheaper model via prompt search.

Getting started

import dspy
dspy.settings.configure(lm=dspy.LM("openai/gpt-4o-mini"))

class ExtractFacts(dspy.Signature):
    """Extract a list of facts from the text."""
    text: str = dspy.InputField()
    facts: list[str] = dspy.OutputField()

extract = dspy.ChainOfThought(ExtractFacts)
print(extract(text="The capital of France is Paris.").facts)

Gotchas

The mental model is different from LangChain. Expect a real ramp-up.
The optimizer can burn a lot of tokens during compilation — use a cheap model for optimization passes.
DSPy’s best results come from iterating on the metric, not the prompt. If your metric is bad, DSPy will cheerfully optimize the wrong thing.