Amazon Bedrock Prompt Management

Amazon Bedrock 提示词管理

Overview

概述

Amazon Bedrock Prompt Management provides enterprise-grade capabilities for creating, versioning, testing, and deploying prompt templates. It enables teams to centralize prompt engineering, implement A/B testing, and integrate prompts across Bedrock Flows, Agents, and applications.

Purpose: Centralized prompt template management with version control, variable substitution, and multi-variant testing

Pattern: Task-based (independent operations for different prompt management tasks)

Key Capabilities:

Prompt Templates - Reusable templates with variable substitution
Version Management - Track changes, rollback, and staged deployment
Multi-Variant Testing - A/B test different prompt variations
Flow Integration - Use prompts in Bedrock Flows and Agents
Variable Types - String, number, array, and JSON object variables
Prompt Catalog - Centralized library for team collaboration
Cross-Model Support - Works with all Bedrock foundation models

Quality Targets:

Reusability: 80%+ prompt template reuse across applications
Version Control: 100% prompt changes tracked
Testing: A/B test 3+ variants per production prompt
Collaboration: Centralized catalog for team-wide access

Amazon Bedrock 提示词管理提供企业级的提示词模板创建、版本控制、测试和部署能力。它支持团队集中进行提示词工程、实施A/B测试，并在Bedrock Flows、Agents和各类应用中集成提示词。

用途：集中管理提示词模板，支持版本控制、变量替换和多变体测试

模式：基于任务（不同提示词管理任务对应独立操作）

核心功能:

提示词模板 - 支持变量替换的可复用模板
版本管理 - 跟踪变更、回滚和分阶段部署
多变体测试 - 对不同提示词变体进行A/B测试
Flow集成 - 在Bedrock Flows和Agents中使用提示词
变量类型 - 支持字符串、数字、数组和JSON对象变量
提示词目录 - 用于团队协作的集中式库
跨模型支持 - 兼容所有Bedrock基础模型

质量目标:

可复用性：跨应用的提示词模板复用率达80%以上
版本控制：100%的提示词变更都被跟踪
测试：每个生产级提示词至少测试3种变体
协作：提供团队全员可访问的集中式目录

When to Use

适用场景

Use bedrock-prompts when:

Creating reusable prompt templates across applications
Managing prompt versions for rollback and staged deployment
Implementing A/B testing for prompt optimization
Building centralized prompt catalogs for teams
Integrating prompts with Bedrock Flows or Agents
Standardizing prompt engineering practices
Testing prompt variations before production deployment
Sharing prompts across multiple projects
Implementing variable substitution in prompts
Optimizing prompts with data-driven testing

When NOT to Use:

Single-use prompts without reuse (use inline prompts)
Simple applications without version control needs
Ad-hoc experimentation (test locally first, then promote to managed prompts)

在以下场景中使用bedrock-prompts：

在多个应用间创建可复用的提示词模板
管理提示词版本以支持回滚和分阶段部署
实施A/B测试来优化提示词
为团队构建集中式提示词目录
将提示词与Bedrock Flows或Agents集成
标准化提示词工程实践
在生产部署前测试提示词变体
在多个项目间共享提示词
在提示词中实现变量替换
通过数据驱动的测试优化提示词

不适用场景:

无需复用的一次性提示词（使用内联提示词）
无需版本控制的简单应用
临时实验（先在本地测试，再推广到托管提示词）

Prerequisites

前置条件

Required

必需条件

AWS account with Bedrock access
IAM permissions for Bedrock Agent service
Foundation model access enabled
boto3 >= 1.34.0

拥有Bedrock访问权限的AWS账户
Bedrock Agent服务的IAM权限
已启用基础模型访问
boto3 >= 1.34.0

Installation

安装

bash

pip install boto3 botocore

bash

pip install boto3 botocore

IAM Permissions

IAM权限

json

{
  "Version": "2012-10-17",
  "Statement": [
    {
      "Effect": "Allow",
      "Action": [
        "bedrock:CreatePrompt",
        "bedrock:GetPrompt",
        "bedrock:UpdatePrompt",
        "bedrock:DeletePrompt",
        "bedrock:ListPrompts",
        "bedrock:CreatePromptVersion",
        "bedrock:ListPromptVersions",
        "bedrock:InvokeModel"
      ],
      "Resource": "*"
    }
  ]
}

json

{
  "Version": "2012-10-17",
  "Statement": [
    {
      "Effect": "Allow",
      "Action": [
        "bedrock:CreatePrompt",
        "bedrock:GetPrompt",
        "bedrock:UpdatePrompt",
        "bedrock:DeletePrompt",
        "bedrock:ListPrompts",
        "bedrock:CreatePromptVersion",
        "bedrock:ListPromptVersions",
        "bedrock:InvokeModel"
      ],
      "Resource": "*"
    }
  ]
}

Quick Start

快速开始

1. Create Prompt Template

1. 创建提示词模板

python

import boto3
import json

bedrock_agent = boto3.client('bedrock-agent', region_name='us-east-1')

response = bedrock_agent.create_prompt(
    name='customer-support-prompt',
    description='Customer support response template',
    variants=[
        {
            'name': 'default',
            'templateType': 'TEXT',
            'modelId': 'anthropic.claude-3-sonnet-20240229-v1:0',
            'templateConfiguration': {
                'text': {
                    'text': '''You are a helpful customer support agent for {{company_name}}.

Customer Query: {{customer_query}}

Instructions:
- Be professional and empathetic
- Provide clear, actionable solutions
- If you don't know, offer to escalate
- Keep responses under {{max_words}} words

Response:''',
                    'inputVariables': [
                        {
                            'name': 'company_name'
                        },
                        {
                            'name': 'customer_query'
                        },
                        {
                            'name': 'max_words'
                        }
                    ]
                }
            },
            'inferenceConfiguration': {
                'text': {
                    'maxTokens': 500,
                    'temperature': 0.7,
                    'topP': 0.9
                }
            }
        }
    ]
)

prompt_id = response['id']
prompt_arn = response['arn']
print(f"Created prompt: {prompt_id}")
print(f"ARN: {prompt_arn}")

python

import boto3
import json

bedrock_agent = boto3.client('bedrock-agent', region_name='us-east-1')

response = bedrock_agent.create_prompt(
    name='customer-support-prompt',
    description='Customer support response template',
    variants=[
        {
            'name': 'default',
            'templateType': 'TEXT',
            'modelId': 'anthropic.claude-3-sonnet-20240229-v1:0',
            'templateConfiguration': {
                'text': {
                    'text': '''You are a helpful customer support agent for {{company_name}}.

Customer Query: {{customer_query}}

Instructions:
- Be professional and empathetic
- Provide clear, actionable solutions
- If you don't know, offer to escalate
- Keep responses under {{max_words}} words

Response:''',
                    'inputVariables': [
                        {
                            'name': 'company_name'
                        },
                        {
                            'name': 'customer_query'
                        },
                        {
                            'name': 'max_words'
                        }
                    ]
                }
            },
            'inferenceConfiguration': {
                'text': {
                    'maxTokens': 500,
                    'temperature': 0.7,
                    'topP': 0.9
                }
            }
        }
    ]
)

prompt_id = response['id']
prompt_arn = response['arn']
print(f"Created prompt: {prompt_id}")
print(f"ARN: {prompt_arn}")

2. Create Prompt Version

2. 创建提示词版本

python

undefined

python

undefined

Create immutable version for production

version_response = bedrock_agent.create_prompt_version( promptIdentifier=prompt_id, description='Production v1.0 - Initial release' )

version = version_response['version'] print(f"Created version: {version}")

undefined

version_response = bedrock_agent.create_prompt_version( promptIdentifier=prompt_id, description='Production v1.0 - Initial release' )

version = version_response['version'] print(f"Created version: {version}")

undefined

3. Get and Use Prompt

3. 获取并使用提示词

python

undefined

python

undefined

Get prompt details

prompt = bedrock_agent.get_prompt( promptIdentifier=prompt_id, promptVersion=version )

Extract template and variables

template = prompt['variants'][0]['templateConfiguration']['text']['text'] variables = {var['name']: None for var in prompt['variants'][0]['templateConfiguration']['text']['inputVariables']}

print(f"Template: {template}") print(f"Variables: {list(variables.keys())}")

---

template = prompt['variants'][0]['templateConfiguration']['text']['text'] variables = {var['name']: None for var in prompt['variants'][0]['templateConfiguration']['text']['inputVariables']}

print(f"Template: {template}") print(f"Variables: {list(variables.keys())}")

---

Operations

操作指南

Operation 1: create-prompt

操作1: create-prompt

Create a new prompt template with variables and inference configuration.

Use when: Building reusable prompt templates, standardizing prompts across applications, creating prompt catalogs

Code Example:

python

import boto3

bedrock_agent = boto3.client('bedrock-agent', region_name='us-east-1')

创建带有变量和推理配置的新提示词模板。

适用场景：构建可复用提示词模板、跨应用标准化提示词、创建提示词目录

代码示例:

python

import boto3

bedrock_agent = boto3.client('bedrock-agent', region_name='us-east-1')

Create advanced prompt with multiple variable types

response = bedrock_agent.create_prompt( name='product-recommendation-prompt', description='E-commerce product recommendation engine', defaultVariant='optimized', variants=[ { 'name': 'optimized', 'templateType': 'TEXT', 'modelId': 'anthropic.claude-3-sonnet-20240229-v1:0', 'templateConfiguration': { 'text': { 'text': '''You are a product recommendation expert for an e-commerce platform.

User Profile:

Name: {{user_name}}
Purchase History: {{purchase_history}}
Preferences: {{preferences}}
Budget: ${{budget}}

Available Categories: {{categories}}

Task: Recommend {{num_recommendations}} products that match the user's profile.

Format your response as a JSON array with product_id, name, price, and reason.''', 'inputVariables': [ {'name': 'user_name'}, {'name': 'purchase_history'}, {'name': 'preferences'}, {'name': 'budget'}, {'name': 'categories'}, {'name': 'num_recommendations'} ] } }, 'inferenceConfiguration': { 'text': { 'maxTokens': 1000, 'temperature': 0.5, 'topP': 0.95, 'stopSequences': ['\n\n---'] } } } ], tags={ 'Environment': 'production', 'Team': 'recommendations', 'CostCenter': 'engineering' } )

print(f"Prompt ID: {response['id']}") print(f"Prompt ARN: {response['arn']}") print(f"Created At: {response['createdAt']}")


**Best Practices**:
- Use descriptive names with hyphens (e.g., `customer-support-prompt`)
- Document variable types and expected formats
- Set appropriate `maxTokens` to control costs
- Use `defaultVariant` to specify preferred version
- Add tags for cost tracking and organization

---

response = bedrock_agent.create_prompt( name='product-recommendation-prompt', description='E-commerce product recommendation engine', defaultVariant='optimized', variants=[ { 'name': 'optimized', 'templateType': 'TEXT', 'modelId': 'anthropic.claude-3-sonnet-20240229-v1:0', 'templateConfiguration': { 'text': { 'text': '''You are a product recommendation expert for an e-commerce platform.

User Profile:

Name: {{user_name}}
Purchase History: {{purchase_history}}
Preferences: {{preferences}}
Budget: ${{budget}}

Available Categories: {{categories}}

Task: Recommend {{num_recommendations}} products that match the user's profile.

Format your response as a JSON array with product_id, name, price, and reason.''', 'inputVariables': [ {'name': 'user_name'}, {'name': 'purchase_history'}, {'name': 'preferences'}, {'name': 'budget'}, {'name': 'categories'}, {'name': 'num_recommendations'} ] } }, 'inferenceConfiguration': { 'text': { 'maxTokens': 1000, 'temperature': 0.5, 'topP': 0.95, 'stopSequences': ['\n\n---'] } } } ], tags={ 'Environment': 'production', 'Team': 'recommendations', 'CostCenter': 'engineering' } )

print(f"Prompt ID: {response['id']}") print(f"Prompt ARN: {response['arn']}") print(f"Created At: {response['createdAt']}")


**最佳实践**:
- 使用带连字符的描述性名称（例如：`customer-support-prompt`）
- 记录变量类型和预期格式
- 设置合适的`maxTokens`以控制成本
- 使用`defaultVariant`指定首选版本
- 添加标签用于成本跟踪和组织管理

---

Operation 2: create-prompt-version

操作2: create-prompt-version

Create immutable versions of prompts for production deployment and rollback.

Use when: Deploying prompts to production, implementing staged rollout, enabling rollback capability

Code Example:

python

import boto3
from datetime import datetime

bedrock_agent = boto3.client('bedrock-agent', region_name='us-east-1')

创建提示词的不可变版本，用于生产部署和回滚。

适用场景：将提示词部署到生产环境、实施分阶段发布、启用回滚能力

代码示例:

python

import boto3
from datetime import datetime

bedrock_agent = boto3.client('bedrock-agent', region_name='us-east-1')

Create version with detailed description

version_response = bedrock_agent.create_prompt_version( promptIdentifier='prompt-12345', description=f'Production v2.0 - {datetime.now().isoformat()} - Added sentiment analysis', tags={ 'Version': '2.0', 'ReleaseDate': datetime.now().strftime('%Y-%m-%d'), 'Changelog': 'Added sentiment context to improve response quality' } )

version_number = version_response['version'] version_arn = version_response['arn']

print(f"Version: {version_number}") print(f"ARN: {version_arn}")

version_response = bedrock_agent.create_prompt_version( promptIdentifier='prompt-12345', description=f'Production v2.0 - {datetime.now().isoformat()} - Added sentiment analysis', tags={ 'Version': '2.0', 'ReleaseDate': datetime.now().strftime('%Y-%m-%d'), 'Changelog': 'Added sentiment context to improve response quality' } )

version_number = version_response['version'] version_arn = version_response['arn']

print(f"Version: {version_number}") print(f"ARN: {version_arn}")

List all versions

list_response = bedrock_agent.list_prompts( promptIdentifier='prompt-12345' )

print("\nAll versions:") for version in list_response.get('promptSummaries', []): print(f"- Version {version['version']}: {version.get('description', 'No description')}")


**Version Management Best Practices**:
- Create versions before production deployment
- Use semantic versioning in descriptions (v1.0, v1.1, v2.0)
- Document changes in version descriptions
- Keep DRAFT version for active development
- Test versions thoroughly before promoting

---

list_response = bedrock_agent.list_prompts( promptIdentifier='prompt-12345' )

print("\nAll versions:") for version in list_response.get('promptSummaries', []): print(f"- Version {version['version']}: {version.get('description', 'No description')}")


**版本管理最佳实践**:
- 在生产部署前创建版本
- 在描述中使用语义化版本（v1.0、v1.1、v2.0）
- 在版本描述中记录变更内容
- 保留DRAFT版本用于活跃开发
- 在推广前彻底测试版本

---

Operation 3: get-prompt

操作3: get-prompt

Retrieve prompt details including template, variables, and inference configuration.

Use when: Inspecting prompt templates, debugging issues, preparing for invocation

Code Example:

python

import boto3

bedrock_agent = boto3.client('bedrock-agent', region_name='us-east-1')

检索提示词详情，包括模板、变量和推理配置。

适用场景：检查提示词模板、调试问题、准备调用

代码示例:

python

import boto3

bedrock_agent = boto3.client('bedrock-agent', region_name='us-east-1')

Get specific version

prompt = bedrock_agent.get_prompt( promptIdentifier='prompt-12345', promptVersion='2' # Omit for DRAFT version )

Extract configuration

variant = prompt['variants'][0] template = variant['templateConfiguration']['text']['text'] variables = variant['templateConfiguration']['text']['inputVariables'] inference_config = variant['inferenceConfiguration']['text']

print(f"Prompt Name: {prompt['name']}") print(f"Version: {prompt['version']}") print(f"Model: {variant['modelId']}") print(f"\nTemplate:\n{template}") print(f"\nVariables:") for var in variables: print(f" - {var['name']}") print(f"\nInference Config:") print(f" Max Tokens: {inference_config['maxTokens']}") print(f" Temperature: {inference_config['temperature']}") print(f" Top P: {inference_config['topP']}")

---

variant = prompt['variants'][0] template = variant['templateConfiguration']['text']['text'] variables = variant['templateConfiguration']['text']['inputVariables'] inference_config = variant['inferenceConfiguration']['text']

print(f"Prompt Name: {prompt['name']}") print(f"Version: {prompt['version']}") print(f"Model: {variant['modelId']}") print(f"\nTemplate:\n{template}") print(f"\nVariables:") for var in variables: print(f" - {var['name']}") print(f"\nInference Config:") print(f" Max Tokens: {inference_config['maxTokens']}") print(f" Temperature: {inference_config['temperature']}") print(f" Top P: {inference_config['topP']}")

---

Operation 4: list-prompts

操作4: list-prompts

List all prompts or filter by criteria.

Use when: Building prompt catalogs, auditing prompt usage, discovering available prompts

Code Example:

python

import boto3

bedrock_agent = boto3.client('bedrock-agent', region_name='us-east-1')

列出所有提示词或按条件筛选。

适用场景：构建提示词目录、审计提示词使用情况、发现可用提示词

代码示例:

python

import boto3

bedrock_agent = boto3.client('bedrock-agent', region_name='us-east-1')

List all prompts with pagination

paginator = bedrock_agent.get_paginator('list_prompts') page_iterator = paginator.paginate()

prompts = [] for page in page_iterator: prompts.extend(page.get('promptSummaries', []))

print(f"Total prompts: {len(prompts)}") print("\nPrompt Catalog:") for prompt in prompts: print(f"\n- {prompt['name']} (ID: {prompt['id']})") print(f" Description: {prompt.get('description', 'N/A')}") print(f" Created: {prompt['createdAt']}") print(f" Updated: {prompt['updatedAt']}") print(f" Version: {prompt['version']}")

---

paginator = bedrock_agent.get_paginator('list_prompts') page_iterator = paginator.paginate()

prompts = [] for page in page_iterator: prompts.extend(page.get('promptSummaries', []))

print(f"Total prompts: {len(prompts)}") print("\nPrompt Catalog:") for prompt in prompts: print(f"\n- {prompt['name']} (ID: {prompt['id']})") print(f" Description: {prompt.get('description', 'N/A')}") print(f" Created: {prompt['createdAt']}") print(f" Updated: {prompt['updatedAt']}") print(f" Version: {prompt['version']}")

---

Operation 5: update-prompt

操作5: update-prompt

Update prompt templates, add variants, or modify inference configuration.

Use when: Improving prompts, adding A/B test variants, adjusting inference parameters

Code Example:

python

import boto3

bedrock_agent = boto3.client('bedrock-agent', region_name='us-east-1')

更新提示词模板、添加变体或修改推理配置。

适用场景：优化提示词、添加A/B测试变体、调整推理参数

代码示例:

python

import boto3

bedrock_agent = boto3.client('bedrock-agent', region_name='us-east-1')

Update prompt with new variant

response = bedrock_agent.update_prompt( promptIdentifier='prompt-12345', name='customer-support-prompt', description='Customer support with multiple response styles', defaultVariant='professional', variants=[ { 'name': 'professional', 'templateType': 'TEXT', 'modelId': 'anthropic.claude-3-sonnet-20240229-v1:0', 'templateConfiguration': { 'text': { 'text': 'Professional tone template...', 'inputVariables': [{'name': 'query'}] } }, 'inferenceConfiguration': { 'text': {'maxTokens': 500, 'temperature': 0.3} } }, { 'name': 'friendly', 'templateType': 'TEXT', 'modelId': 'anthropic.claude-3-sonnet-20240229-v1:0', 'templateConfiguration': { 'text': { 'text': 'Friendly tone template...', 'inputVariables': [{'name': 'query'}] } }, 'inferenceConfiguration': { 'text': {'maxTokens': 500, 'temperature': 0.7} } } ] )

print(f"Updated prompt: {response['id']}")

---

response = bedrock_agent.update_prompt( promptIdentifier='prompt-12345', name='customer-support-prompt', description='Customer support with multiple response styles', defaultVariant='professional', variants=[ { 'name': 'professional', 'templateType': 'TEXT', 'modelId': 'anthropic.claude-3-sonnet-20240229-v1:0', 'templateConfiguration': { 'text': { 'text': 'Professional tone template...', 'inputVariables': [{'name': 'query'}] } }, 'inferenceConfiguration': { 'text': {'maxTokens': 500, 'temperature': 0.3} } }, { 'name': 'friendly', 'templateType': 'TEXT', 'modelId': 'anthropic.claude-3-sonnet-20240229-v1:0', 'templateConfiguration': { 'text': { 'text': 'Friendly tone template...', 'inputVariables': [{'name': 'query'}] } }, 'inferenceConfiguration': { 'text': {'maxTokens': 500, 'temperature': 0.7} } } ] )

print(f"Updated prompt: {response['id']}")

---

Operation 6: delete-prompt

操作6: delete-prompt

Delete prompt templates (cannot be undone).

Use when: Cleaning up unused prompts, removing deprecated templates

Code Example:

python

import boto3

bedrock_agent = boto3.client('bedrock-agent', region_name='us-east-1')

删除提示词模板（不可撤销）。

适用场景：清理未使用的提示词、移除已弃用的模板

代码示例:

python

import boto3

bedrock_agent = boto3.client('bedrock-agent', region_name='us-east-1')

Delete prompt (all versions)

response = bedrock_agent.delete_prompt( promptIdentifier='prompt-12345' )

print(f"Deleted prompt: {response['id']}") print(f"Status: {response['status']}")


**Warning**: Deletion is permanent and affects all versions. Ensure prompt is not used in Flows or Agents before deleting.

---

response = bedrock_agent.delete_prompt( promptIdentifier='prompt-12345' )

print(f"Deleted prompt: {response['id']}") print(f"Status: {response['status']}")


**警告**：删除操作是永久性的，会影响所有版本。删除前请确保提示词未在Flows或Agents中使用。

---

Variable Types and Substitution

变量类型与替换

Supported Variable Types

支持的变量类型

Bedrock Prompt Management supports multiple variable types:

String: Text values (default)
Number: Numeric values
Array: Lists of items
JSON Object: Complex structured data

Bedrock 提示词管理支持多种变量类型：

字符串：文本值（默认类型）
数字：数值类型
数组：项目列表
JSON对象：复杂结构化数据

Variable Substitution Example

变量替换示例

python

import boto3
import json

bedrock_agent = boto3.client('bedrock-agent', region_name='us-east-1')
bedrock_runtime = boto3.client('bedrock-runtime', region_name='us-east-1')

python

import boto3
import json

bedrock_agent = boto3.client('bedrock-agent', region_name='us-east-1')
bedrock_runtime = boto3.client('bedrock-runtime', region_name='us-east-1')

Create prompt with complex variables

prompt_response = bedrock_agent.create_prompt( name='data-analysis-prompt', variants=[{ 'name': 'default', 'templateType': 'TEXT', 'modelId': 'anthropic.claude-3-sonnet-20240229-v1:0', 'templateConfiguration': { 'text': { 'text': '''Analyze the following data:

Dataset: {{dataset_name}} Columns: {{columns}} Row Count: {{row_count}} Sample Data: {{sample_data}}

Analysis Type: {{analysis_type}}

Provide insights and recommendations.''', 'inputVariables': [ {'name': 'dataset_name'}, {'name': 'columns'}, {'name': 'row_count'}, {'name': 'sample_data'}, {'name': 'analysis_type'} ] } } }] )

prompt_response = bedrock_agent.create_prompt( name='data-analysis-prompt', variants=[{ 'name': 'default', 'templateType': 'TEXT', 'modelId': 'anthropic.claude-3-sonnet-20240229-v1:0', 'templateConfiguration': { 'text': { 'text': '''Analyze the following data:

Dataset: {{dataset_name}} Columns: {{columns}} Row Count: {{row_count}} Sample Data: {{sample_data}}

Analysis Type: {{analysis_type}}

Provide insights and recommendations.''', 'inputVariables': [ {'name': 'dataset_name'}, {'name': 'columns'}, {'name': 'row_count'}, {'name': 'sample_data'}, {'name': 'analysis_type'} ] } } }] )

Use prompt with variable substitution

prompt = bedrock_agent.get_prompt( promptIdentifier=prompt_response['id'] )

template = prompt['variants'][0]['templateConfiguration']['text']['text']

prompt = bedrock_agent.get_prompt( promptIdentifier=prompt_response['id'] )

template = prompt['variants'][0]['templateConfiguration']['text']['text']

Substitute variables

variables = { 'dataset_name': 'Sales Q4 2024', 'columns': json.dumps(['date', 'product', 'revenue', 'quantity']), 'row_count': '10,000', 'sample_data': json.dumps([ {'date': '2024-10-01', 'product': 'Widget A', 'revenue': 1500, 'quantity': 50}, {'date': '2024-10-02', 'product': 'Widget B', 'revenue': 2000, 'quantity': 75} ]), 'analysis_type': 'Revenue trends and product performance' }

Replace variables in template

prompt_text = template for var_name, var_value in variables.items(): prompt_text = prompt_text.replace(f'{{{{{var_name}}}}}', str(var_value))

print(f"Final Prompt:\n{prompt_text}")

---

prompt_text = template for var_name, var_value in variables.items(): prompt_text = prompt_text.replace(f'{{{{{var_name}}}}}', str(var_value))

print(f"Final Prompt:\n{prompt_text}")

---

Multi-Variant Testing (A/B Testing)

多变体测试（A/B测试）

Creating Multi-Variant Prompts

创建多变体提示词

python

import boto3

bedrock_agent = boto3.client('bedrock-agent', region_name='us-east-1')

python

import boto3

bedrock_agent = boto3.client('bedrock-agent', region_name='us-east-1')

Create prompt with 3 variants for A/B/C testing

response = bedrock_agent.create_prompt( name='email-subject-generator', description='A/B/C test for email subject lines', defaultVariant='variant-a', variants=[ { 'name': 'variant-a', 'templateType': 'TEXT', 'modelId': 'anthropic.claude-3-sonnet-20240229-v1:0', 'templateConfiguration': { 'text': { 'text': 'Generate a professional email subject line for: {{email_content}}', 'inputVariables': [{'name': 'email_content'}] } }, 'inferenceConfiguration': { 'text': {'maxTokens': 50, 'temperature': 0.3} } }, { 'name': 'variant-b', 'templateType': 'TEXT', 'modelId': 'anthropic.claude-3-sonnet-20240229-v1:0', 'templateConfiguration': { 'text': { 'text': 'Create an engaging, click-worthy subject line for: {{email_content}}', 'inputVariables': [{'name': 'email_content'}] } }, 'inferenceConfiguration': { 'text': {'maxTokens': 50, 'temperature': 0.7} } }, { 'name': 'variant-c', 'templateType': 'TEXT', 'modelId': 'anthropic.claude-3-haiku-20240307-v1:0', 'templateConfiguration': { 'text': { 'text': 'Write a concise, action-oriented subject line for: {{email_content}}', 'inputVariables': [{'name': 'email_content'}] } }, 'inferenceConfiguration': { 'text': {'maxTokens': 50, 'temperature': 0.5} } } ] )

print(f"Created multi-variant prompt: {response['id']}")

undefined

response = bedrock_agent.create_prompt( name='email-subject-generator', description='A/B/C test for email subject lines', defaultVariant='variant-a', variants=[ { 'name': 'variant-a', 'templateType': 'TEXT', 'modelId': 'anthropic.claude-3-sonnet-20240229-v1:0', 'templateConfiguration': { 'text': { 'text': 'Generate a professional email subject line for: {{email_content}}', 'inputVariables': [{'name': 'email_content'}] } }, 'inferenceConfiguration': { 'text': {'maxTokens': 50, 'temperature': 0.3} } }, { 'name': 'variant-b', 'templateType': 'TEXT', 'modelId': 'anthropic.claude-3-sonnet-20240229-v1:0', 'templateConfiguration': { 'text': { 'text': 'Create an engaging, click-worthy subject line for: {{email_content}}', 'inputVariables': [{'name': 'email_content'}] } }, 'inferenceConfiguration': { 'text': {'maxTokens': 50, 'temperature': 0.7} } }, { 'name': 'variant-c', 'templateType': 'TEXT', 'modelId': 'anthropic.claude-3-haiku-20240307-v1:0', 'templateConfiguration': { 'text': { 'text': 'Write a concise, action-oriented subject line for: {{email_content}}', 'inputVariables': [{'name': 'email_content'}] } }, 'inferenceConfiguration': { 'text': {'maxTokens': 50, 'temperature': 0.5} } } ] )

print(f"Created multi-variant prompt: {response['id']}")

undefined

Testing Framework

测试框架

python

import boto3
import random
import json
from datetime import datetime

class PromptABTester:
    def __init__(self, prompt_id, region='us-east-1'):
        self.bedrock_agent = boto3.client('bedrock-agent', region_name=region)
        self.bedrock_runtime = boto3.client('bedrock-runtime', region_name=region)
        self.prompt_id = prompt_id
        self.results = []

    def get_variants(self):
        prompt = self.bedrock_agent.get_prompt(promptIdentifier=self.prompt_id)
        return [v['name'] for v in prompt['variants']]

    def test_variant(self, variant_name, variables, user_id=None):
        # Get prompt
        prompt = self.bedrock_agent.get_prompt(promptIdentifier=self.prompt_id)

        # Find variant
        variant = next(v for v in prompt['variants'] if v['name'] == variant_name)

        # Substitute variables
        template = variant['templateConfiguration']['text']['text']
        for var_name, var_value in variables.items():
            template = template.replace(f'{{{{{var_name}}}}}', str(var_value))

        # Invoke model
        model_id = variant['modelId']
        inference_config = variant['inferenceConfiguration']['text']

        response = self.bedrock_runtime.invoke_model(
            modelId=model_id,
            body=json.dumps({
                'anthropic_version': 'bedrock-2023-05-31',
                'messages': [{'role': 'user', 'content': template}],
                'max_tokens': inference_config['maxTokens'],
                'temperature': inference_config['temperature']
            })
        )

        result = json.loads(response['body'].read())
        output = result['content'][0]['text']

        # Record test result
        self.results.append({
            'timestamp': datetime.now().isoformat(),
            'variant': variant_name,
            'user_id': user_id,
            'input_variables': variables,
            'output': output,
            'model': model_id
        })

        return output

    def random_test(self, variables, user_id=None):
        """Randomly select variant for testing"""
        variants = self.get_variants()
        selected_variant = random.choice(variants)
        return self.test_variant(selected_variant, variables, user_id)

    def analyze_results(self):
        """Analyze test results by variant"""
        analysis = {}
        for result in self.results:
            variant = result['variant']
            if variant not in analysis:
                analysis[variant] = {
                    'count': 0,
                    'avg_output_length': 0,
                    'samples': []
                }

            analysis[variant]['count'] += 1
            analysis[variant]['samples'].append(result['output'])
            analysis[variant]['avg_output_length'] += len(result['output'])

        # Calculate averages
        for variant in analysis:
            count = analysis[variant]['count']
            analysis[variant]['avg_output_length'] /= count

        return analysis

python

import boto3
import random
import json
from datetime import datetime

class PromptABTester:
    def __init__(self, prompt_id, region='us-east-1'):
        self.bedrock_agent = boto3.client('bedrock-agent', region_name=region)
        self.bedrock_runtime = boto3.client('bedrock-runtime', region_name=region)
        self.prompt_id = prompt_id
        self.results = []

    def get_variants(self):
        prompt = self.bedrock_agent.get_prompt(promptIdentifier=self.prompt_id)
        return [v['name'] for v in prompt['variants']]

    def test_variant(self, variant_name, variables, user_id=None):
        # Get prompt
        prompt = self.bedrock_agent.get_prompt(promptIdentifier=self.prompt_id)

        # Find variant
        variant = next(v for v in prompt['variants'] if v['name'] == variant_name)

        # Substitute variables
        template = variant['templateConfiguration']['text']['text']
        for var_name, var_value in variables.items():
            template = template.replace(f'{{{{{var_name}}}}}', str(var_value))

        # Invoke model
        model_id = variant['modelId']
        inference_config = variant['inferenceConfiguration']['text']

        response = self.bedrock_runtime.invoke_model(
            modelId=model_id,
            body=json.dumps({
                'anthropic_version': 'bedrock-2023-05-31',
                'messages': [{'role': 'user', 'content': template}],
                'max_tokens': inference_config['maxTokens'],
                'temperature': inference_config['temperature']
            })
        )

        result = json.loads(response['body'].read())
        output = result['content'][0]['text']

        # Record test result
        self.results.append({
            'timestamp': datetime.now().isoformat(),
            'variant': variant_name,
            'user_id': user_id,
            'input_variables': variables,
            'output': output,
            'model': model_id
        })

        return output

    def random_test(self, variables, user_id=None):
        """Randomly select variant for testing"""
        variants = self.get_variants()
        selected_variant = random.choice(variants)
        return self.test_variant(selected_variant, variables, user_id)

    def analyze_results(self):
        """Analyze test results by variant"""
        analysis = {}
        for result in self.results:
            variant = result['variant']
            if variant not in analysis:
                analysis[variant] = {
                    'count': 0,
                    'avg_output_length': 0,
                    'samples': []
                }

            analysis[variant]['count'] += 1
            analysis[variant]['samples'].append(result['output'])
            analysis[variant]['avg_output_length'] += len(result['output'])

        # Calculate averages
        for variant in analysis:
            count = analysis[variant]['count']
            analysis[variant]['avg_output_length'] /= count

        return analysis

Usage

tester = PromptABTester('prompt-12345')

Run A/B test

for i in range(100): result = tester.random_test( variables={'email_content': f'Test email content {i}'}, user_id=f'user-{i}' )

Analyze

analysis = tester.analyze_results() print(json.dumps(analysis, indent=2))

---

analysis = tester.analyze_results() print(json.dumps(analysis, indent=2))

---

Prompt Engineering Best Practices

提示词工程最佳实践

1. Clear Instructions

1. 清晰的指令

python

undefined

python

undefined

Good: Specific, clear instructions

prompt = '''You are a financial analyst.

Task: Analyze the following quarterly earnings data and provide:

Revenue trends (% change YoY)
Key growth drivers
Risk factors

Data: {{financial_data}}

Format: Use bullet points. Keep analysis under 200 words.'''

prompt = '''You are a financial analyst.

Task: Analyze the following quarterly earnings data and provide:

Revenue trends (% change YoY)
Key growth drivers
Risk factors

Data: {{financial_data}}

Format: Use bullet points. Keep analysis under 200 words.'''

Bad: Vague, unclear

prompt = '''Analyze this: {{financial_data}}'''

undefined

prompt = '''Analyze this: {{financial_data}}'''

undefined

2. Variable Naming Conventions

2. 变量命名规范

python

undefined

python

undefined

Good: Descriptive variable names

inputVariables=[ {'name': 'customer_query'}, {'name': 'customer_purchase_history'}, {'name': 'max_response_words'} ]

Bad: Ambiguous names

inputVariables=[ {'name': 'input'}, {'name': 'data'}, {'name': 'limit'} ]

undefined

inputVariables=[ {'name': 'input'}, {'name': 'data'}, {'name': 'limit'} ]

undefined

3. Inference Configuration

3. 推理配置

python

undefined

python

undefined

Creative tasks: Higher temperature

'inferenceConfiguration': { 'text': { 'maxTokens': 1000, 'temperature': 0.8, # More creative 'topP': 0.95 } }

Factual tasks: Lower temperature

'inferenceConfiguration': { 'text': { 'maxTokens': 500, 'temperature': 0.1, # More deterministic 'topP': 0.9 } }

undefined

'inferenceConfiguration': { 'text': { 'maxTokens': 500, 'temperature': 0.1, # More deterministic 'topP': 0.9 } }

undefined

4. Stop Sequences

4. 停止序列

python

undefined

python

undefined

Use stop sequences to control output length

'inferenceConfiguration': { 'text': { 'maxTokens': 1000, 'stopSequences': ['\n\n---', 'END_RESPONSE', '###'] } }

---

'inferenceConfiguration': { 'text': { 'maxTokens': 1000, 'stopSequences': ['\n\n---', 'END_RESPONSE', '###'] } }

---

Integration with Bedrock Flows

与Bedrock Flows集成

Using Prompts in Flows

在Flows中使用提示词

python

import boto3

bedrock_agent = boto3.client('bedrock-agent', region_name='us-east-1')

python

import boto3

bedrock_agent = boto3.client('bedrock-agent', region_name='us-east-1')

Create flow that uses managed prompt

flow_response = bedrock_agent.create_flow( name='customer-support-flow', executionRoleArn='arn:aws:iam::123456789012:role/BedrockFlowRole', definition={ 'nodes': [ { 'name': 'FlowInput', 'type': 'Input', 'outputs': [{'name': 'query', 'type': 'String'}] }, { 'name': 'SupportPrompt', 'type': 'Prompt', 'configuration': { 'prompt': { 'sourceConfiguration': { 'resource': { 'promptArn': 'arn:aws:bedrock:us-east-1:123456789012:prompt/prompt-12345:2' } } } }, 'inputs': [ { 'name': 'customer_query', 'expression': 'FlowInput.query' }, { 'name': 'company_name', 'expression': '"Acme Corp"' }, { 'name': 'max_words', 'expression': '150' } ], 'outputs': [{'name': 'response', 'type': 'String'}] }, { 'name': 'FlowOutput', 'type': 'Output', 'inputs': [ { 'name': 'response', 'expression': 'SupportPrompt.response' } ] } ], 'connections': [ {'source': 'FlowInput', 'target': 'SupportPrompt'}, {'source': 'SupportPrompt', 'target': 'FlowOutput'} ] } )

print(f"Created flow: {flow_response['id']}")

---

flow_response = bedrock_agent.create_flow( name='customer-support-flow', executionRoleArn='arn:aws:iam::123456789012:role/BedrockFlowRole', definition={ 'nodes': [ { 'name': 'FlowInput', 'type': 'Input', 'outputs': [{'name': 'query', 'type': 'String'}] }, { 'name': 'SupportPrompt', 'type': 'Prompt', 'configuration': { 'prompt': { 'sourceConfiguration': { 'resource': { 'promptArn': 'arn:aws:bedrock:us-east-1:123456789012:prompt/prompt-12345:2' } } } }, 'inputs': [ { 'name': 'customer_query', 'expression': 'FlowInput.query' }, { 'name': 'company_name', 'expression': '"Acme Corp"' }, { 'name': 'max_words', 'expression': '150' } ], 'outputs': [{'name': 'response', 'type': 'String'}] }, { 'name': 'FlowOutput', 'type': 'Output', 'inputs': [ { 'name': 'response', 'expression': 'SupportPrompt.response' } ] } ], 'connections': [ {'source': 'FlowInput', 'target': 'SupportPrompt'}, {'source': 'SupportPrompt', 'target': 'FlowOutput'} ] } )

print(f"Created flow: {flow_response['id']}")

---

Related Skills

Complete Example: Production Prompt Catalog

完整示例：生产级提示词目录

python

import boto3
import json
from typing import Dict, List, Optional

class PromptCatalog:
    """Enterprise prompt catalog with versioning and testing"""

    def __init__(self, region='us-east-1'):
        self.bedrock_agent = boto3.client('bedrock-agent', region_name=region)
        self.bedrock_runtime = boto3.client('bedrock-runtime', region_name=region)
        self.catalog = {}

    def create_prompt_template(
        self,
        name: str,
        description: str,
        template: str,
        variables: List[str],
        model_id: str = 'anthropic.claude-3-sonnet-20240229-v1:0',
        max_tokens: int = 1000,
        temperature: float = 0.7,
        tags: Optional[Dict] = None
    ) -> str:
        """Create a new prompt template"""

        response = self.bedrock_agent.create_prompt(
            name=name,
            description=description,
            variants=[{
                'name': 'default',
                'templateType': 'TEXT',
                'modelId': model_id,
                'templateConfiguration': {
                    'text': {
                        'text': template,
                        'inputVariables': [{'name': var} for var in variables]
                    }
                },
                'inferenceConfiguration': {
                    'text': {
                        'maxTokens': max_tokens,
                        'temperature': temperature,
                        'topP': 0.9
                    }
                }
            }],
            tags=tags or {}
        )

        prompt_id = response['id']
        self.catalog[name] = prompt_id
        return prompt_id

    def version_prompt(self, name: str, description: str) -> str:
        """Create immutable version"""
        prompt_id = self.catalog[name]
        response = self.bedrock_agent.create_prompt_version(
            promptIdentifier=prompt_id,
            description=description
        )
        return response['version']

    def get_prompt(self, name: str, version: Optional[str] = None) -> Dict:
        """Get prompt by name"""
        prompt_id = self.catalog[name]
        return self.bedrock_agent.get_prompt(
            promptIdentifier=prompt_id,
            promptVersion=version
        ) if version else self.bedrock_agent.get_prompt(promptIdentifier=prompt_id)

    def list_catalog(self) -> List[Dict]:
        """List all prompts in catalog"""
        response = self.bedrock_agent.list_prompts()
        return response.get('promptSummaries', [])

python

import boto3
import json
from typing import Dict, List, Optional

class PromptCatalog:
    """Enterprise prompt catalog with versioning and testing"""

    def __init__(self, region='us-east-1'):
        self.bedrock_agent = boto3.client('bedrock-agent', region_name=region)
        self.bedrock_runtime = boto3.client('bedrock-runtime', region_name=region)
        self.catalog = {}

    def create_prompt_template(
        self,
        name: str,
        description: str,
        template: str,
        variables: List[str],
        model_id: str = 'anthropic.claude-3-sonnet-20240229-v1:0',
        max_tokens: int = 1000,
        temperature: float = 0.7,
        tags: Optional[Dict] = None
    ) -> str:
        """Create a new prompt template"""

        response = self.bedrock_agent.create_prompt(
            name=name,
            description=description,
            variants=[{
                'name': 'default',
                'templateType': 'TEXT',
                'modelId': model_id,
                'templateConfiguration': {
                    'text': {
                        'text': template,
                        'inputVariables': [{'name': var} for var in variables]
                    }
                },
                'inferenceConfiguration': {
                    'text': {
                        'maxTokens': max_tokens,
                        'temperature': temperature,
                        'topP': 0.9
                    }
                }
            }],
            tags=tags or {}
        )

        prompt_id = response['id']
        self.catalog[name] = prompt_id
        return prompt_id

    def version_prompt(self, name: str, description: str) -> str:
        """Create immutable version"""
        prompt_id = self.catalog[name]
        response = self.bedrock_agent.create_prompt_version(
            promptIdentifier=prompt_id,
            description=description
        )
        return response['version']

    def get_prompt(self, name: str, version: Optional[str] = None) -> Dict:
        """Get prompt by name"""
        prompt_id = self.catalog[name]
        return self.bedrock_agent.get_prompt(
            promptIdentifier=prompt_id,
            promptVersion=version
        ) if version else self.bedrock_agent.get_prompt(promptIdentifier=prompt_id)

    def list_catalog(self) -> List[Dict]:
        """List all prompts in catalog"""
        response = self.bedrock_agent.list_prompts()
        return response.get('promptSummaries', [])

Usage

catalog = PromptCatalog(region='us-east-1')

Create customer support prompt

support_id = catalog.create_prompt_template( name='customer-support-v1', description='Customer support response generator', template='''You are a customer support agent for {{company_name}}.

Query: {{query}}

Provide a helpful, professional response in under {{max_words}} words.''', variables=['company_name', 'query', 'max_words'], max_tokens=500, temperature=0.5, tags={'Department': 'Support', 'Environment': 'Production'} )

support_id = catalog.create_prompt_template( name='customer-support-v1', description='Customer support response generator', template='''You are a customer support agent for {{company_name}}.

Query: {{query}}

Provide a helpful, professional response in under {{max_words}} words.''', variables=['company_name', 'query', 'max_words'], max_tokens=500, temperature=0.5, tags={'Department': 'Support', 'Environment': 'Production'} )

Create version

version = catalog.version_prompt('customer-support-v1', 'Initial production release')

print(f"Created prompt: {support_id}") print(f"Version: {version}")

version = catalog.version_prompt('customer-support-v1', 'Initial production release')

print(f"Created prompt: {support_id}") print(f"Version: {version}")

List catalog

prompts = catalog.list_catalog() print(f"\nCatalog contains {len(prompts)} prompts")

---

prompts = catalog.list_catalog() print(f"\nCatalog contains {len(prompts)} prompts")

---

Summary

总结

Amazon Bedrock Prompt Management provides enterprise-grade prompt template capabilities:

Centralized Management: Single source of truth for prompts
Version Control: Track changes, rollback, staged deployment
A/B Testing: Multi-variant testing for optimization
Variable Substitution: Flexible templating system
Flow Integration: Use prompts across Bedrock services
Team Collaboration: Shared prompt catalog
Production Ready: Immutable versions, tagging, monitoring

Use bedrock-prompts to standardize prompt engineering, enable A/B testing, and build reusable prompt libraries for production AI applications.

Amazon Bedrock 提示词管理提供企业级的提示词模板能力：

集中管理：提示词的单一可信来源
版本控制：跟踪变更、回滚和分阶段部署
A/B测试：多变体测试以优化效果
变量替换：灵活的模板系统
Flow集成：在Bedrock服务间复用提示词
团队协作：共享的提示词目录
生产就绪：不可变版本、标签、监控

使用bedrock-prompts来标准化提示词工程、启用A/B测试，并为生产级AI应用构建可复用的提示词库。

bedrock-prompts

Original

Translation

Amazon Bedrock Prompt Management

Amazon Bedrock 提示词管理

Overview

概述

When to Use

适用场景

Prerequisites

前置条件

Required

必需条件

Recommended

推荐条件

Installation

安装

IAM Permissions

IAM权限

Quick Start

快速开始

1. Create Prompt Template

1. 创建提示词模板

2. Create Prompt Version

2. 创建提示词版本

Create immutable version for production

Create immutable version for production

3. Get and Use Prompt

3. 获取并使用提示词

Get prompt details

Get prompt details

Extract template and variables

Extract template and variables

Operations

操作指南

Operation 1: create-prompt

操作1: create-prompt

Create advanced prompt with multiple variable types

Create advanced prompt with multiple variable types

Operation 2: create-prompt-version

操作2: create-prompt-version

Create version with detailed description

Create version with detailed description

List all versions

List all versions

Operation 3: get-prompt

操作3: get-prompt

Get specific version

Get specific version

Extract configuration

Extract configuration

Operation 4: list-prompts

操作4: list-prompts

List all prompts with pagination

List all prompts with pagination

Operation 5: update-prompt

操作5: update-prompt

Update prompt with new variant

Update prompt with new variant

Operation 6: delete-prompt

操作6: delete-prompt

Delete prompt (all versions)

Delete prompt (all versions)

Variable Types and Substitution

变量类型与替换

Supported Variable Types

支持的变量类型

Variable Substitution Example

变量替换示例

Create prompt with complex variables

Create prompt with complex variables

Use prompt with variable substitution

Use prompt with variable substitution

Substitute variables

Substitute variables

Replace variables in template

Replace variables in template

Multi-Variant Testing (A/B Testing)

多变体测试（A/B测试）

Creating Multi-Variant Prompts