首页

主题

Prompts ​

1. Prompt Templates (提示模板) ​

基础模板 ​

python
from langchain.prompts import PromptTemplate

# 基础模板示例
template = """
问题: {question}
请一步一步回答以上的问题
"""

prompt = PromptTemplate(
    input_variables=["question"],
    template=template
)

# 使用模板
formatted_prompt = prompt.format(
    question="江苏的省会是哪里"
)

print(formatted_prompt)

输出结果

shell
问题: 江苏的省会是哪里
请一步一步回答以上的问题

聊天模板 ​

python
from langchain.prompts import ChatPromptTemplate
from langchain.prompts.chat import SystemMessagePromptTemplate, HumanMessagePromptTemplate

# 系统消息模板
system_template = "You are a helpful assistant that translates {input_language} to {output_language}."
system_prompt = SystemMessagePromptTemplate.from_template(system_template)

# 人类消息模板
human_template = "{text}"
human_prompt = HumanMessagePromptTemplate.from_template(human_template)

# 组合聊天模板
chat_prompt = ChatPromptTemplate.from_messages([
    system_prompt,
    human_prompt
])

# 使用模板
messages = chat_prompt.format_messages(
    input_language="English",
    output_language="French",
    text="I love programming."
)
print(messages)

输出结果

shell
[SystemMessage(content='You are a helpful assistant that translates English to French.', additional_kwargs={}, response_metadata={}), HumanMessage(content='I love programming.', additional_kwargs={}, response_metadata={})]

高级特性 ​

1. 部分变量填充 ​

python
from langchain.prompts import PromptTemplate

prompt = PromptTemplate(
    template="You are a {role}. Help {name} with {task}.",
    input_variables=["name", "task", "role"]
)

# 部分填充
partial_prompt = prompt.partial(role="professional programmer")
# 后续可以只填充其他变量
final_prompt = partial_prompt.format(
    name="Patrick",
    task="debugging Python code"
)
print(final_prompt)

输出

shell
You are a professional programmer. Help Patrick with debugging Python code.

2. 模板验证 ​

python
from langchain.prompts import PromptTemplate
from pydantic import BaseModel, validator

class CustomPromptTemplate(PromptTemplate):
    @validator('input_variables')
    def validate_variables(cls, v):
        if 'required_field' not in v:
            raise ValueError("required_field must be an input variable")
        return v

2. Example Selectors (示例选择器) ​

类型 ​

  1. LengthBasedExampleSelector (基于长度选择)
  2. SemanticSimilarityExampleSelector (基于语义相似度)
  3. NGramOverlapExampleSelector (基于N-gram重叠)

实现示例 ​

python
from langchain.prompts.example_selector import SemanticSimilarityExampleSelector
from langchain.vectorstores import Chroma
from langchain.embeddings import OpenAIEmbeddings

# 准备示例
examples = [
    {"input": "What is the capital of France?", "output": "The capital of France is Paris."},
    {"input": "What is the largest planet?", "output": "Jupiter is the largest planet."},
    {"input": "Who wrote Romeo and Juliet?", "output": "William Shakespeare wrote Romeo and Juliet."}
]

# 创建选择器
example_selector = SemanticSimilarityExampleSelector.from_examples(
    examples,
    OpenAIEmbeddings(),
    Chroma,
    k=2  # 返回最相似的2个示例
)

# 使用选择器
selected_examples = example_selector.select_examples({"input": "What is the capital of Spain?"})

3. Output Parsers (输出解析器) ​

常用解析器 ​

  1. PydanticOutputParser
  2. CommaSeparatedListOutputParser
  3. StructuredOutputParser
  4. RegexParser

实现示例 ​

Pydantic 解析器 ​

python
from langchain.output_parsers import PydanticOutputParser
from pydantic import BaseModel, Field
from typing import List

class Analysis(BaseModel):
    summary: str = Field(description="Brief summary of the main points")
    keywords: List[str] = Field(description="Key terms or concepts")
    sentiment: str = Field(description="Overall sentiment (positive/negative/neutral)")

parser = PydanticOutputParser(pydantic_object=Analysis)

prompt = PromptTemplate(
    template="Analyze the following text:\n{text}\n{format_instructions}",
    input_variables=["text"],
    partial_variables={"format_instructions": parser.get_format_instructions()}
)

结构化输出解析器 ​

python
from langchain.output_parsers import StructuredOutputParser, ResponseSchema

response_schemas = [
    ResponseSchema(name="answer", description="The final answer to the question"),
    ResponseSchema(name="confidence", description="Confidence score from 0 to 100"),
    ResponseSchema(name="reasoning", description="The step-by-step reasoning process")
]

parser = StructuredOutputParser.from_response_schemas(response_schemas)

prompt = PromptTemplate(
    template="Question: {question}\n{format_instructions}\n",
    input_variables=["question"],
    partial_variables={"format_instructions": parser.get_format_instructions()}
)

错误处理 ​

python
from langchain.output_parsers import OutputFixingParser

# 创建修复解析器
fixing_parser = OutputFixingParser.from_llm(
    parser=parser,
    llm=llm
)

try:
    parsed_output = parser.parse(llm_output)
except:
    # 尝试修复并重新解析
    parsed_output = fixing_parser.parse(llm_output)

最佳实践 ​

  1. 模板设计

    • 使用清晰的指令
    • 提供具体的示例
    • 指定输出格式
    • 包含错误处理指南

  2. 示例选择

    • 保持示例多样性
    • 根据任务特点选择合适的选择器
    • 定期更新示例库

  3. 输出解析

    • 使用适当的验证规则
    • 实现优雅的错误处理
    • 考虑输出修复机制

  4. 性能优化

    • 缓存常用提示
    • 批量处理示例
    • 异步处理大量请求