Skip to main content

Dspy

DSPy:声明式语言模型程序,自动优化提示,RAG。

技能元数据

来源内置(默认安装)
路径skills/mlops/research/dspy
版本1.0.0
作者Orchestra Research
许可证MIT
依赖项dspy, openai, anthropic
平台linux, macos, windows
标签提示工程, DSPy, 声明式编程, RAG, Agents, 提示优化, LM 编程, 斯坦福 NLP, 自动优化, 模块化 AI

参考:完整 SKILL.md

info

以下是 Hermes 在触发此技能时加载的完整技能定义。当技能激活时,Agent 会看到这些指令。

DSPy:声明式语言模型编程

何时使用此技能

在以下情况下使用 DSPy:

  • 构建复杂的 AI 系统,包含多个组件和工作流
  • 以声明式方式编程 LM,而非手动提示工程
  • 使用数据驱动方法自动优化提示
  • 创建模块化 AI 流水线,易于维护和移植
  • 使用优化器系统性地改进模型输出
  • 构建 RAG 系统、Agents 或分类器,获得更高可靠性

GitHub Stars:22,000+ | 创建者:斯坦福 NLP

安装

# 稳定版
pip install dspy

# 最新开发版
pip install git+https://github.com/stanfordnlp/dspy.git

# 使用特定 LM 提供商
pip install dspy[openai] # OpenAI
pip install dspy[anthropic] # Anthropic Claude
pip install dspy[all] # 所有提供商

快速开始

基本示例:问答

import dspy

# 配置你的语言模型
lm = dspy.Claude(model="claude-sonnet-4-5-20250929")
dspy.settings.configure(lm=lm)

# 定义签名(输入 → 输出)
class QA(dspy.Signature):
"""用简短的事实性答案回答问题。"""
question = dspy.InputField()
answer = dspy.OutputField(desc="通常为 1 到 5 个词")

# 创建模块
qa = dspy.Predict(QA)

# 使用它
response = qa(question="法国的首都是什么?")
print(response.answer) # "巴黎"

思维链推理

import dspy

lm = dspy.Claude(model="claude-sonnet-4-5-20250929")
dspy.settings.configure(lm=lm)

# 使用 ChainOfThought 获得更好的推理
class MathProblem(dspy.Signature):
"""解决数学文字题。"""
problem = dspy.InputField()
answer = dspy.OutputField(desc="数值答案")

# ChainOfThought 自动生成推理步骤
cot = dspy.ChainOfThought(MathProblem)

response = cot(problem="如果约翰有 5 个苹果,给了玛丽 2 个,他还剩几个?")
print(response.rationale) # 显示推理步骤
print(response.answer) # "3"

核心概念

1. 签名

签名定义了 AI 任务的结构(输入 → 输出):

# 内联签名(简单)
qa = dspy.Predict("question -> answer")

# 类签名(详细)
class Summarize(dspy.Signature):
"""将文本总结为关键点。"""
text = dspy.InputField()
summary = dspy.OutputField(desc="要点,3-5 项")

summarizer = dspy.ChainOfThought(Summarize)

何时使用:

  • 内联:快速原型开发,简单任务
  • :复杂任务,类型提示,更好的文档

2. 模块

模块是将输入转换为输出的可复用组件:

dspy.Predict

基础预测模块:

predictor = dspy.Predict("context, question -> answer")
result = predictor(context="巴黎是法国的首都",
question="首都是什么?")

dspy.ChainOfThought

在回答前生成推理步骤:

cot = dspy.ChainOfThought("question -> answer")
result = cot(question="为什么天空是蓝色的?")
print(result.rationale) # 推理步骤
print(result.answer) # 最终答案

dspy.ReAct

类似 Agent 的推理,支持工具:

from dspy.predict import ReAct

class SearchQA(dspy.Signature):
"""使用搜索回答问题。"""
question = dspy.InputField()
answer = dspy.OutputField()

def search_tool(query: str) -> str:
"""搜索维基百科。"""
# 你的搜索实现
return results

react = ReAct(SearchQA, tools=[search_tool])
result = react(question="Python 是什么时候创建的?")

dspy.ProgramOfThought

生成并执行代码进行推理:

pot = dspy.ProgramOfThought("question -> answer")
result = pot(question="240 的 15% 是多少?")
# 生成:answer = 240 * 0.15

3. 优化器

优化器使用训练数据自动改进你的模块:

BootstrapFewShot

从示例中学习:

from dspy.teleprompt import BootstrapFewShot

# 训练数据
trainset = [
dspy.Example(question="2+2 等于多少?", answer="4").with_inputs("question"),
dspy.Example(question="3+5 等于多少?", answer="8").with_inputs("question"),
]

# 定义指标
def validate_answer(example, pred, trace=None):
return example.answer == pred.answer

# 优化
optimizer = BootstrapFewShot(metric=validate_answer, max_bootstrapped_demos=3)
optimized_qa = optimizer.compile(qa, trainset=trainset)

# 现在 optimized_qa 表现更好!

MIPRO(最重要提示优化)

迭代改进提示:

from dspy.teleprompt import MIPRO

optimizer = MIPRO(
metric=validate_answer,
num_candidates=10,
init_temperature=1.0
)

optimized_cot = optimizer.compile(
cot,
trainset=trainset,
num_trials=100
)

BootstrapFinetune

为模型微调创建数据集:

from dspy.teleprompt import BootstrapFinetune

optimizer = BootstrapFinetune(metric=validate_answer)
optimized_module = optimizer.compile(qa, trainset=trainset)

# 导出用于微调的训练数据

4. 构建复杂系统

多阶段流水线

import dspy

class MultiHopQA(dspy.Module):
def __init__(self):
super().__init__()
self.retrieve = dspy.Retrieve(k=3)
self.generate_query = dspy.ChainOfThought("question -> search_query")
self.generate_answer = dspy.ChainOfThought("context, question -> answer")

def forward(self, question):
# Stage 1: Generate search query
search_query = self.generate_query(question=question).search_query

# Stage 2: Retrieve context
passages = self.retrieve(search_query).passages
context = "\n".join(passages)

# Stage 3: Generate answer
answer = self.generate_answer(context=context, question=question).answer
return dspy.Prediction(answer=answer, context=context)

# Use the pipeline
qa_system = MultiHopQA()
result = qa_system(question="Who wrote the book that inspired the movie Blade Runner?")

带优化的 RAG 系统

import dspy
from dspy.retrieve.chromadb_rm import ChromadbRM

# Configure retriever
retriever = ChromadbRM(
collection_name="documents",
persist_directory="./chroma_db"
)

class RAG(dspy.Module):
def __init__(self, num_passages=3):
super().__init__()
self.retrieve = dspy.Retrieve(k=num_passages)
self.generate = dspy.ChainOfThought("context, question -> answer")

def forward(self, question):
context = self.retrieve(question).passages
return self.generate(context=context, question=question)

# Create and optimize
rag = RAG()

# Optimize with training data
from dspy.teleprompt import BootstrapFewShot

optimizer = BootstrapFewShot(metric=validate_answer)
optimized_rag = optimizer.compile(rag, trainset=trainset)

语言模型提供商配置

Anthropic Claude

import dspy

lm = dspy.Claude(
model="claude-sonnet-4-5-20250929",
api_key="your-api-key", # Or set ANTHROPIC_API_KEY env var
max_tokens=1000,
temperature=0.7
)
dspy.settings.configure(lm=lm)

OpenAI

lm = dspy.OpenAI(
model="gpt-4",
api_key="your-api-key",
max_tokens=1000
)
dspy.settings.configure(lm=lm)

本地模型 (Ollama)

lm = dspy.OllamaLocal(
model="llama3.1",
base_url="http://localhost:11434"
)
dspy.settings.configure(lm=lm)

多模型

# Different models for different tasks
cheap_lm = dspy.OpenAI(model="gpt-3.5-turbo")
strong_lm = dspy.Claude(model="claude-sonnet-4-5-20250929")

# Use cheap model for retrieval, strong model for reasoning
with dspy.settings.context(lm=cheap_lm):
context = retriever(question)

with dspy.settings.context(lm=strong_lm):
answer = generator(context=context, question=question)

常见模式

模式 1:结构化输出

from pydantic import BaseModel, Field

class PersonInfo(BaseModel):
name: str = Field(description="Full name")
age: int = Field(description="Age in years")
occupation: str = Field(description="Current job")

class ExtractPerson(dspy.Signature):
"""Extract person information from text."""
text = dspy.InputField()
person: PersonInfo = dspy.OutputField()

extractor = dspy.TypedPredictor(ExtractPerson)
result = extractor(text="John Doe is a 35-year-old software engineer.")
print(result.person.name) # "John Doe"
print(result.person.age) # 35

Pattern 2: 断言驱动优化

import dspy
from dspy.primitives.assertions import assert_transform_module, backtrack_handler

class MathQA(dspy.Module):
def __init__(self):
super().__init__()
self.solve = dspy.ChainOfThought("problem -> solution: float")

def forward(self, problem):
solution = self.solve(problem=problem).solution

# 断言解决方案是数字
dspy.Assert(
isinstance(float(solution), float),
"Solution must be a number",
backtrack=backtrack_handler
)

return dspy.Prediction(solution=solution)

Pattern 3: 自一致性

import dspy
from collections import Counter

class ConsistentQA(dspy.Module):
def __init__(self, num_samples=5):
super().__init__()
self.qa = dspy.ChainOfThought("question -> answer")
self.num_samples = num_samples

def forward(self, question):
# 生成多个答案
answers = []
for _ in range(self.num_samples):
result = self.qa(question=question)
answers.append(result.answer)

# 返回最常见的答案
most_common = Counter(answers).most_common(1)[0][0]
return dspy.Prediction(answer=most_common)

Pattern 4: 带重排序的检索

class RerankedRAG(dspy.Module):
def __init__(self):
super().__init__()
self.retrieve = dspy.Retrieve(k=10)
self.rerank = dspy.Predict("question, passage -> relevance_score: float")
self.answer = dspy.ChainOfThought("context, question -> answer")

def forward(self, question):
# 检索候选段落
passages = self.retrieve(question).passages

# 对段落重排序
scored = []
for passage in passages:
score = float(self.rerank(question=question, passage=passage).relevance_score)
scored.append((score, passage))

# 取前 3 个
top_passages = [p for _, p in sorted(scored, reverse=True)[:3]]
context = "\n\n".join(top_passages)

# 生成答案
return self.answer(context=context, question=question)

评估与指标

自定义指标

def exact_match(example, pred, trace=None):
"""精确匹配指标。"""
return example.answer.lower() == pred.answer.lower()

def f1_score(example, pred, trace=None):
"""文本重叠的 F1 分数。"""
pred_tokens = set(pred.answer.lower().split())
gold_tokens = set(example.answer.lower().split())

if not pred_tokens:
return 0.0

precision = len(pred_tokens & gold_tokens) / len(pred_tokens)
recall = len(pred_tokens & gold_tokens) / len(gold_tokens)

if precision + recall == 0:
return 0.0

return 2 * (precision * recall) / (precision + recall)

评估

from dspy.evaluate import Evaluate

# 创建评估器
evaluator = Evaluate(
devset=testset,
metric=exact_match,
num_threads=4,
display_progress=True
)

# 评估模型
score = evaluator(qa_system)
print(f"Accuracy: {score}")

# 比较优化前后的模型
score_before = evaluator(qa)
score_after = evaluator(optimized_qa)
print(f"Improvement: {score_after - score_before:.2%}")

最佳实践

1. 从简单开始,逐步迭代

# Start with Predict
qa = dspy.Predict("question -> answer")

# Add reasoning if needed
qa = dspy.ChainOfThought("question -> answer")

# Add optimization when you have data
optimized_qa = optimizer.compile(qa, trainset=data)

2. 使用描述性的 Signature

# ❌ Bad: Vague
class Task(dspy.Signature):
input = dspy.InputField()
output = dspy.OutputField()

# ✅ Good: Descriptive
class SummarizeArticle(dspy.Signature):
"""Summarize news articles into 3-5 key points."""
article = dspy.InputField(desc="full article text")
summary = dspy.OutputField(desc="bullet points, 3-5 items")

3. 使用代表性数据优化

# Create diverse training examples
trainset = [
dspy.Example(question="factual", answer="...).with_inputs("question"),
dspy.Example(question="reasoning", answer="...").with_inputs("question"),
dspy.Example(question="calculation", answer="...").with_inputs("question"),
]

# Use validation set for metric
def metric(example, pred, trace=None):
return example.answer in pred.answer

4. 保存和加载优化后的模型

# Save
optimized_qa.save("models/qa_v1.json")

# Load
loaded_qa = dspy.ChainOfThought("question -> answer")
loaded_qa.load("models/qa_v1.json")

5. 监控与调试

# Enable tracing
dspy.settings.configure(lm=lm, trace=[])

# Run prediction
result = qa(question="...")

# Inspect trace
for call in dspy.settings.trace:
print(f"Prompt: {call['prompt']}")
print(f"Response: {call['response']}")

与其他方法的比较

特性手动提示LangChainDSPy
提示工程手动手动自动
优化试错数据驱动
模块化
类型安全有限是(Signature)
可移植性
学习曲线中高

何时选择 DSPy:

  • 你有训练数据或能生成它
  • 你需要系统地改进提示
  • 你在构建复杂的多阶段系统
  • 你想跨不同语言模型进行优化

何时选择替代方案:

  • 快速原型(手动提示)
  • 使用现有工具的简单链(LangChain)
  • 需要自定义优化逻辑

资源

参见

  • references/modules.md - 模块详细指南(Predict、ChainOfThought、ReAct、ProgramOfThought)
  • references/optimizers.md - 优化算法(BootstrapFewShot、MIPRO、BootstrapFinetune)
  • references/examples.md - 实际示例(RAG、Agents、分类器)