Program, don’t prompt,your LLMs.

DSPy 3.4.0 — PythonInterpreter improvements, faster GEPA, and MCP v2 compatibility · learn more →

Program, don’t prompt,

your LLMs.

DSPy is a Python framework for building AI systems. Express your tasks as structured signatures, not prompts, to produce maintainable, modular, and optimizable programs.

$ pip install -U dspy Getting Started →

python ≥ 3.10 MIT license Stanford NLP github.com/stanfordnlp/dspy

Change LLM Add a Field Make it an Agent

extract\_events.py

12345 678910

lm = dspy.LM("openai/gpt-5.4-nano")

class ExtractEvent(dspy.Signature):

"""Extract event details from an email."""

email: str \= dspy.InputField()

event\_name: str \= dspy.OutputField()

date: str \= dspy.OutputField()

extract = dspy.Predict(ExtractEvent)

extract(email=inbox\_message)

Prediction(

event\_name="Team Offsite",

date="Thursday, June 5"

)

5.2M+

monthly downloads

461+

contributors

38k

github stars

in production at

Databricks Shopify Dropbox

Compose programs with reusable primitives.

Signatures

Declare your task.

Define your task as typed inputs and outputs instead of managing messy prompts. Portable, maintainable, and easy to iterate on.

Learn about Signatures →

class Triage(dspy.Signature):

"""Route a support ticket."""

ticket: str \= dspy.InputField()

urgency: Literal\["low", "high"\] = dspy.OutputField()

team: str \= dspy.OutputField()

Modules

Same interface, different strategy.

Modules control how your signature executes. Reason, run ensembles, use tools, add a REPL, and more without rewriting your task.

Explore Modules →

\# Direct completion

classify = dspy.Predict(Triage)

\# Add step-by-step reasoning

classify = dspy.ChainOfThought(Triage)

\# Add tools and a reasoning loop

classify = dspy.ReAct(Triage, tools=\[search\])

Optimizers

Compile your program against a metric.

Give DSPy examples and a scoring function. It tunes your prompts automatically until quality converges.

Try Optimizers →

tp = dspy.GEPA(

metric=semantic\_f1,

auto="medium")

opt = tp.compile(rag, trainset)

\# Before: 0.41 F1

\# After: 0.63 F1

opt.save("rag.v2.json")

Extract Agent Pipeline Multimodal Optimize

class Extract(dspy.Signature):

"""Extract contact info."""

message: str \= dspy.InputField()

name: str \= dspy.OutputField()

email: Optional\[str\] = dspy.OutputField()

intent: Literal\[

"meeting", "intro", "follow-up"

\] = dspy.OutputField()

extract = dspy.Predict(Extract)

extract(message="I'm Sarah"

"(sarah@acme.co). Meet Thursday?")

Prediction(

name="Sarah",

email="sarah@acme.co",

intent="meeting"

)

def search(query: str) -> list\[str\]:

"""Search a knowledge base."""

return kb.query(query, k=3)

def calc(expr: str) -> float:

"""Evaluate a math expression."""

return dspy.PythonInterpreter({}).execute(expr)

agent = dspy.ReAct(

"question -> answer",

tools=\[search, calc\])

agent(question="GDP per capita of France?")

\# thought 1: I need France's GDP and population.

\# action 1: search("France GDP") → ...

\# thought 2: Now divide GDP by population.

\# action 2: calc("3.13e12 / 68e6") → 46029.4

Prediction(answer="$46,029")

class FactCheck(dspy.Module):

def \\_init\\_(self):

self.find = dspy.ChainOfThought(

"article -> claims: list\[str\]")

self.verify = dspy.ChainOfThought(

"claim, source -> verdict")

def forward(self, article):

found = self.find(article=article)

return \[

self.verify(claim=c, source=article)

for c in found.claims\]

\# >>> FactCheck()(article=news\_article)

\[Prediction(verdict="supported"),

Prediction(verdict="unsupported"),

Prediction(verdict="supported")\]

class AnalyzeChart(dspy.Signature):

"""Describe the trend and key data points in a chart."""

chart: dspy.Image \= dspy.InputField()

title: str \= dspy.OutputField()

trend: str \= dspy.OutputField()

data\_points: list\[dict\] = dspy.OutputField()

analyze = dspy.Predict(AnalyzeChart)

analyze(chart=dspy.Image("quarterly\_revenue.png"))

Prediction(

title="Quarterly Revenue (2024)",

trend="Steady growth, Q3 dip, strong Q4 recovery",

data\_points=\[{"q": "Q1", "rev": "$4.2M"}, ...\]

)

optimizer = dspy.GEPA(

metric=accuracy, auto="medium")

optimized = optimizer.compile(

extract, trainset=labeled\_emails)

optimized.save("extract\_v2.json")

\# Baseline 62% (gpt-5.4-mini, zero-shot)

\# Optimized 89% (gpt-5.4-mini + GEPA compile)

\# Cost $2.18 · 200 examples

\# Saved to → extract\_v2.json

Built in the open, since Dec 2022.

DSPy started at Stanford NLP and grew into a research community. New optimizers and module types land here first — then show up in production systems at companies you’ve heard of.

Dec 2025

Recursive Language Models

arxiv →

Jul 2025

GEPA: Reflective Prompt Evolution

arxiv →

Jul 2024

BetterTogether: Fine-Tuning + Prompt Opt.

arxiv →

Jun 2024

MIPROv2: Optimizing Instructions & Demos

arxiv →