提示优化 API 参考¶

函数

create_prompt_optimizer –

创建一个提示优化器，用于提高提示的有效性。
create_multi_prompt_optimizer –

创建一个多提示优化器，用于提高提示的有效性。

create_prompt_optimizer ¶

create_prompt_optimizer(
    model: str | BaseChatModel,
    /,
    *,
    kind: KINDS = "gradient",
    config: Union[
        GradientOptimizerConfig,
        MetapromptOptimizerConfig,
        None,
    ] = None,
) -> Runnable[OptimizerInput, str]

创建一个提示优化器，用于提高提示的有效性。

此函数创建一个优化器，可以分析和改进提示，以提高其在语言模型中的表现。它支持多种优化策略，以迭代方式增强提示的质量和有效性。

参数

model (Union[str, BaseChatModel]) –

用于优化的语言模型。可以是模型名称字符串或 BaseChatModel 实例。
kind (Literal[gradient, prompt_memory, metaprompt], default: 'gradient' ) –
要使用的优化策略。每种策略都提供不同的优势
- gradient: 分离查找改进领域和推荐更新的关注点
- prompt_memory: 简单的单次元提示
- metaprompt: 支持反思，但每一步都是一次 LLM 调用。

config (Optional[OptimizerConfig], default: None ) –

优化器的配置选项。类型取决于所选策略

- GradientOptimizerConfig for kind="gradient"
- PromptMemoryConfig for kind="prompt_memory"
- MetapromptOptimizerConfig for kind="metaprompt"

默认为 None。

返回值

optimizer ( Runnable[OptimizerInput, str] ) –

一个可调用对象，接受对话轨迹和/或提示并返回优化后的版本。

优化策略¶

1. 梯度优化器¶

sequenceDiagram
    participant U as User
    participant O as Optimizer
    participant R as Reflection
    participant U2 as Update

    U->>O: Prompt + Feedback
    loop For min_steps to max_steps
        O->>R: Think/Critique Current State
        R-->>O: Proposed Improvements
        O->>U2: Apply Update
        U2-->>O: Updated Prompt
    end
    O->>U: Final Optimized Prompt

梯度优化器使用反思来提出改进建议

通过反思周期分析提示和反馈
提出具体的改进建议
应用单步更新

配置 (GradientOptimizerConfig)

gradient_prompt: 用于预测“要改进什么”的自定义提示
metaprompt: 用于应用改进的自定义提示
max_reflection_steps: 最大反思迭代次数（默认值：3）
min_reflection_steps: 最小反思迭代次数（默认值：1）

2. 元提示优化器¶

sequenceDiagram
    participant U as User
    participant M as MetaOptimizer
    participant A as Analysis
    participant U2 as Update

    U->>M: Prompt + Examples
    M->>A: Analyze Examples
    A-->>M: Proposed Update
    M->>U2: Apply Update
    U2-->>U: Enhanced Prompt

使用元学习直接提出更新建议

分析示例以理解模式
提出直接的提示更新
在单个步骤中应用更新

配置 (MetapromptOptimizerConfig)

metaprompt: 关于如何更新提示的自定义指令
max_reflection_steps: 最大元学习步数（默认值：3）
min_reflection_steps: 最小元学习步数（默认值：1）

3. 提示记忆优化器¶

sequenceDiagram
    participant U as User
    participant P as PromptMemory
    participant M as Memory

    U->>P: Prompt + History
    P->>M: Extract Patterns
    M-->>P: Success Patterns
    P->>U: Updated Prompt

从对话历史中学习

从过去的交互中提取成功模式
从反馈中识别改进领域
将学习到的模式应用于新提示

无需额外配置。

示例

基本提示优化

from langmem import create_prompt_optimizer

optimizer = create_prompt_optimizer("anthropic:claude-3-5-sonnet-latest")

# Example conversation with feedback
conversation = [
    {"role": "user", "content": "Tell me about the solar system"},
    {"role": "assistant", "content": "The solar system consists of..."},
]
feedback = {"clarity": "needs more structure"}

# Use conversation history to improve the prompt
trajectories = [(conversation, feedback)]
better_prompt = await optimizer.ainvoke(
    {"trajectories": trajectories, "prompt": "You are an astronomy expert"}
)
print(better_prompt)
# Output: 'Provide a comprehensive overview of the solar system...'

使用对话反馈进行优化

from langmem import create_prompt_optimizer

optimizer = create_prompt_optimizer(
    "anthropic:claude-3-5-sonnet-latest", kind="prompt_memory"
)

# Conversation with feedback about what could be improved
conversation = [
    {"role": "user", "content": "How do I write a bash script?"},
    {"role": "assistant", "content": "Let me explain bash scripting..."},
]
feedback = "Response should include a code example"

# Use the conversation and feedback to improve the prompt
trajectories = [(conversation, {"feedback": feedback})]
better_prompt = await optimizer(trajectories, "You are a coding assistant")
print(better_prompt)
# Output: 'You are a coding assistant that always includes...'

针对复杂任务的元提示优化

from langmem import create_prompt_optimizer

optimizer = create_prompt_optimizer(
    "anthropic:claude-3-5-sonnet-latest",
    kind="metaprompt",
    config={"max_reflection_steps": 3, "min_reflection_steps": 1},
)

# Complex conversation that needs better structure
conversation = [
    {"role": "user", "content": "Explain quantum computing"},
    {"role": "assistant", "content": "Quantum computing uses..."},
]
feedback = "Need better organization and concrete examples"

# Optimize with meta-learning
trajectories = [(conversation, feedback)]
improved_prompt = await optimizer(
    trajectories, "You are a quantum computing expert"
)

性能考量

每种策略有不同的 LLM 调用模式

prompt_memory: 总共 1 次 LLM 调用
- 最快，因为它只需要一次通过
metaprompt: 1-5 次 LLM 调用（可配置）
- 每一步是一次 LLM 调用
- 默认范围：最小 2，最大 5 反思步
gradient: 2-10 次 LLM 调用（可配置）
- 每一步需要 2 次 LLM 调用（思考 + 评论）
- 默认范围：最小 2，最大 5 反思步

策略选择

根据您的需求选择

提示记忆：最简单的提示策略
- 从复杂模式中学习的能力有限
元提示：速度与改进的平衡
- 适中成本（2-5 次 LLM 调用）
梯度：最彻底但昂贵
- 最高成本（4-10 次 LLM 调用）
- 使用关注点分离从更具对话性的上下文中提取反馈。

create_multi_prompt_optimizer ¶

create_multi_prompt_optimizer(
    model: str | BaseChatModel,
    /,
    *,
    kind: Literal[
        "gradient", "prompt_memory", "metaprompt"
    ] = "gradient",
    config: Optional[dict] = None,
) -> Runnable[MultiPromptOptimizerInput, list[Prompt]]

创建一个多提示优化器，用于提高提示的有效性。

此函数创建一个优化器，可以使用相同的优化策略同时分析和改进多个提示。每个提示都使用选定的策略进行优化（有关策略详情，请参阅 create_prompt_optimizer）。

参数

model (Union[str, BaseChatModel]) –

用于优化的语言模型。可以是模型名称字符串或 BaseChatModel 实例。
kind (Literal[gradient, prompt_memory, metaprompt], default: 'gradient' ) –

要使用的优化策略。每种策略都提供不同的优势：- gradient：通过反思迭代改进 - prompt_memory：使用成功的过去提示 - metaprompt：通过元学习学习最优模式默认为“gradient”。
config (Optional[OptimizerConfig], default: None ) –

优化器的配置选项。类型取决于所选策略：- 当 kind="gradient" 时使用 GradientOptimizerConfig - 当 kind="prompt_memory" 时使用 PromptMemoryConfig - 当 kind="metaprompt" 时使用 MetapromptOptimizerConfig 默认为 None。

返回值

MultiPromptOptimizer ( Runnable[MultiPromptOptimizerInput, list[Prompt]] ) –

一个 Runnable，接受对话轨迹和提示并返回优化后的版本。

sequenceDiagram
    participant U as User
    participant M as Multi-prompt Optimizer
    participant C as Credit Assigner
    participant O as Single-prompt Optimizer
    participant P as Prompts

    U->>M: Annotated Trajectories + Prompts
    activate M
    Note over M: Using pre-initialized<br/>single-prompt optimizer

    M->>C: Analyze trajectories
    activate C
    Note over C: Determine which prompts<br/>need improvement
    C-->>M: Credit assignment results
    deactivate C

    loop For each prompt needing update
        M->>O: Optimize prompt
        activate O
        O->>P: Apply optimization strategy
        Note over O,P: Gradient/Memory/Meta<br/>optimization
        P-->>O: Optimized prompt
        O-->>M: Return result
        deactivate O
    end

    M->>U: Return optimized prompts
    deactivate M

系统优化器

示例

基本提示优化

from langmem import create_multi_prompt_optimizer

optimizer = create_multi_prompt_optimizer("anthropic:claude-3-5-sonnet-latest")

# Example conversation with feedback
conversation = [
    {"role": "user", "content": "Tell me about the solar system"},
    {"role": "assistant", "content": "The solar system consists of..."},
]
feedback = {"clarity": "needs more structure"}

# Use conversation history to improve the prompts
trajectories = [(conversation, feedback)]
prompts = [
    {"name": "research", "prompt": "Research the given topic thoroughly"},
    {"name": "summarize", "prompt": "Summarize the research findings"},
]
better_prompts = await optimizer.ainvoke(
    {"trajectories": trajectories, "prompts": prompts}
)
print(better_prompts)

使用对话反馈进行优化

from langmem import create_multi_prompt_optimizer

optimizer = create_multi_prompt_optimizer(
    "anthropic:claude-3-5-sonnet-latest", kind="prompt_memory"
)

# Conversation with feedback about what could be improved
conversation = [
    {"role": "user", "content": "How do I write a bash script?"},
    {"role": "assistant", "content": "Let me explain bash scripting..."},
]
feedback = "Response should include a code example"

# Use the conversation and feedback to improve the prompts
trajectories = [(conversation, {"feedback": feedback})]
prompts = [
    {"name": "explain", "prompt": "Explain the concept"},
    {"name": "example", "prompt": "Provide a practical example"},
]
better_prompts = await optimizer(trajectories, prompts)

控制最大反思步数

from langmem import create_multi_prompt_optimizer

optimizer = create_multi_prompt_optimizer(
    "anthropic:claude-3-5-sonnet-latest",
    kind="metaprompt",
    config={"max_reflection_steps": 3, "min_reflection_steps": 1},
)

# Complex conversation that needs better structure
conversation = [
    {"role": "user", "content": "Explain quantum computing"},
    {"role": "assistant", "content": "Quantum computing uses..."},
]
# Explicit feedback is optional
feedback = None

# Optimize with meta-learning
trajectories = [(conversation, feedback)]
prompts = [
    {"name": "concept", "prompt": "Explain quantum concepts"},
    {"name": "application", "prompt": "Show practical applications"},
    {"name": "example", "prompt": "Give concrete examples"},
]
improved_prompts = await optimizer(trajectories, prompts)

类

Prompt –

用于结构化提示管理和优化的 TypedDict。
OptimizerInput –

单提示优化的输入。
MultiPromptOptimizerInput –

用于一起优化多个提示并保持一致性的输入。
AnnotatedTrajectory –

带有可选反馈的对话历史记录（消息列表），用于提示优化。

Prompt ¶

Bases: TypedDict

用于结构化提示管理和优化的 TypedDict。

示例

from langmem import Prompt

prompt = Prompt(
    name="extract_entities",
    prompt="Extract key entities from the text:",
    update_instructions="Make minimal changes, only address where"
    " errors have occurred after reasoning over why they occur.",
    when_to_update="If there seem to be errors in recall of named entities.",
)

name 和 prompt 字段是必需的。可选字段控制优化：- update_instructions：修改提示的指南 - when_to_update：优化期间提示之间的依赖关系

在提示优化器中使用。

OptimizerInput ¶

Bases: TypedDict

单提示优化的输入。

示例

{
    "trajectories": [
        AnnotatedTrajectory(
            messages=[
                {"role": "user", "content": "What's the weather like?"},
                {
                    "role": "assistant",
                    "content": "I'm sorry, I can't tell you that",
                },
            ],
            feedback="Should have checked your search tool.",
        ),
    ],
    "prompt": Prompt(
        name="main_assistant",
        prompt="You are a helpful assistant with a search tool.",
        update_instructions="Make minimal changes, only address where "
        "errors have occurred after reasoning over why they occur.",
        when_to_update="Any time you notice the agent behaving in a way that doesn't help the user.",
    ),
}

MultiPromptOptimizerInput ¶

Bases: TypedDict

用于一起优化多个提示并保持一致性的输入。

示例

{
    "trajectories": [
        AnnotatedTrajectory(
            messages=[
                {"role": "user", "content": "Tell me about this image"},
                {
                    "role": "assistant",
                    "content": "I see a dog playing in a park",
                },
                {"role": "user", "content": "What breed is it?"},
                {
                    "role": "assistant",
                    "content": "Sorry, I can't tell the breed",
                },
            ],
            feedback="Vision model wasn't used for breed detection",
        ),
    ],
    "prompts": [
        Prompt(
            name="vision_extract",
            prompt="Extract visual details from the image",
            update_instructions="Focus on using vision model capabilities",
        ),
        Prompt(
            name="vision_classify",
            prompt="Classify specific attributes in the image",
            when_to_update="After vision_extract is updated",
        ),
    ],
}

AnnotatedTrajectory ¶

Bases: NamedTuple

带有可选反馈的对话历史记录（消息列表），用于提示优化。

示例

from langmem.prompts.types import AnnotatedTrajectory

trajectory = AnnotatedTrajectory(
    messages=[
        {"role": "user", "content": "What pizza is good around here?"},
        {"role": "assistant", "content": "Try LangPizza™️"},
        {"role": "user", "content": "Stop advertising to me."},
        {"role": "assistant", "content": "BUT YOU'LL LOVE IT!"},
    ],
    feedback={
        "developer_feedback": "too pushy",
        "score": 0,
    },
)