最后更新于
prompts/
├── production/ # 生产环境（软链接指向具体版本）
│ ├── customer_service.yaml → ../versions/customer_service/v2.1.0.yaml
│ └── content_moderation.yaml → ../versions/content_moderation/v1.3.0.yaml
│
├── versions/ # 历史版本
│ ├── customer_service/
│ │ ├── v1.0.0.yaml
│ │ ├── v2.0.0.yaml
│ │ └── v2.1.0.yaml
│ └── content_moderation/
│ ├── v1.0.0.yaml
│ └── v1.3.0.yaml
│
├── development/ # 开发中版本
│ └── customer_service_v3.yaml
│
└── tests/ # 测试用例
├── customer_service/
│ └── test_cases.yaml
└── content_moderation/
└── test_cases.yamlMAJOR.MINOR.PATCH
MAJOR: 不兼容的重大变更（如输出格式改变）
MINOR: 向后兼容的功能改进（如新增处理能力）
PATCH: 向后兼容的问题修复（如措辞优化）
# customer_service_v2.1.0.yaml
metadata:
name: customer_service
version: 2.1.0
description: 客服对话提示词
author: prompt_team
created_at: 2025-01-10
last_updated: 2025-01-14
model_compatibility:
- gpt-4-turbo
- claude-sonnet-4-6
config:
model: gpt-4-turbo
temperature: 0.3
max_tokens: 1000
prompts:
system: |
你是[公司名称]的智能客服助手。
核心职责：
1. 回答产品和服务相关问题
2. 处理订单查询和退换货
3. 收集用户反馈
行为准则：
- 保持专业、友好的语气
- 如遇无法处理的问题，引导用户联系人工客服
- 不透露内部流程和系统信息
user_template: |
用户历史：{user_history}
当前问题：{user_message}
tags:
- customer_service
- production
changelog:
- version: 2.1.0
date: 2025-01-14
changes:
- 优化了退换货流程引导
- 增加了情感识别响应
- version: 2.0.0
date: 2025-01-01
changes:
- 重构提示词结构
- 采用 XML 标签分隔
# environments.yaml
environments:
development:
description: 开发环境，快速迭代
model: gpt-3.5-turbo # 降低成本
rate_limit: 100/min
logging: verbose
staging:
description: 预发布环境，全面测试
model: gpt-4-turbo
rate_limit: 500/min
logging: standard
production:
description: 生产环境，稳定运行
model: gpt-4-turbo
rate_limit: 1000/min
logging: minimal
fallback_model: gpt-3.5-turboclass PromptManager:
def __init__(self, environment="development"):
self.env = environment
self.config = load_env_config(environment)
def get_prompt(self, prompt_name):
if self.env == "production":
return load_prompt(f"production/{prompt_name}.yaml")
elif self.env == "staging":
return load_prompt(f"staging/{prompt_name}.yaml")
else:
return load_prompt(f"development/{prompt_name}.yaml")
def deploy_to_staging(self, prompt_name, version):
"""将指定版本部署到预发布环境"""
source = f"versions/{prompt_name}/{version}.yaml"
target = f"staging/{prompt_name}.yaml"
symlink(source, target)
def deploy_to_production(self, prompt_name, from_staging=True):
"""将预发布版本提升到生产环境"""
if from_staging:
source = f"staging/{prompt_name}.yaml"
target = f"production/{prompt_name}.yaml"
symlink(source, target)
log_deployment(prompt_name, self.get_version(source))
# monitoring_config.yaml
metrics:
performance:
- name: response_latency_p50
description: 响应延迟中位数
threshold: 2000ms
- name: response_latency_p99
description: 响应延迟 99 分位
threshold: 5000ms
quality:
- name: user_satisfaction_rate
description: 用户满意度评分
threshold: 4.0/5.0
- name: escalation_rate
description: 转人工率
threshold: 15%
- name: rejection_rate
description: 安全拒绝率
threshold: 5%
cost:
- name: tokens_per_conversation
description: 每次对话 token 消耗
threshold: 2000
- name: daily_cost
description: 日成本
threshold: $500
alerts:
- condition: response_latency_p99 > 5000ms for 5min
severity: warning
notify: [slack_channel]
- condition: rejection_rate > 10% for 10min
severity: critical
notify: [slack_channel, pagerduty]
- condition: daily_cost > $1000
severity: warning
notify: [email_team, slack_channel]class GradualRollout:
def __init__(self, old_version, new_version):
self.old_version = old_version
self.new_version = new_version
self.new_version_percentage = 0
def get_prompt(self, request_id):
"""根据流量百分比返回对应版本"""
if hash(request_id) % 100 < self.new_version_percentage:
return self.new_version
return self.old_version
def increase_traffic(self, percentage):
"""增加新版本流量占比"""
self.new_version_percentage = min(100, self.new_version_percentage + percentage)
log(f"New version traffic: {self.new_version_percentage}%")
def rollback(self):
"""回滚到旧版本"""
self.new_version_percentage = 0
log("Rolled back to old version")阶段 1: 1% 流量 (金丝雀)
↓ 监控 1 小时，确认无异常
阶段 2: 10% 流量
↓ 监控 4 小时，对比新旧版本指标
阶段 3: 50% 流量
↓ 监控 24 小时，收集用户反馈
阶段 4: 100% 流量
↓ 持续监控，准备快速回滚def auto_rollout_decision(metrics, thresholds):
"""基于指标自动决策是否继续灰度"""
if metrics.error_rate > thresholds.error_rate_max:
return "ROLLBACK", "Error rate exceeded threshold"
if metrics.latency_p99 > thresholds.latency_max:
return "ROLLBACK", "Latency exceeded threshold"
if metrics.user_satisfaction < thresholds.satisfaction_min:
return "HOLD", "User satisfaction below threshold"
if metrics.sample_size < thresholds.min_samples:
return "HOLD", "Insufficient sample size"
return "PROCEED", "All metrics within acceptable range"class RollbackManager:
def __init__(self):
self.deployment_history = []
def deploy(self, prompt_name, version):
"""部署新版本并记录历史"""
previous = get_current_version(prompt_name)
self.deployment_history.append({
"prompt": prompt_name,
"from_version": previous,
"to_version": version,
"timestamp": now(),
"deployed_by": current_user()
})
set_production_version(prompt_name, version)
def rollback(self, prompt_name, steps=1):
"""回滚到之前的版本"""
history = [h for h in self.deployment_history if h["prompt"] == prompt_name]
if len(history) < steps:
raise Error("Insufficient history for rollback")
target = history[-steps]["from_version"]
set_production_version(prompt_name, target)
log(f"Rolled back {prompt_name} to {target}")
def emergency_rollback(self, prompt_name):
"""紧急回滚到最后已知稳定版本"""
stable = get_last_stable_version(prompt_name)
set_production_version(prompt_name, stable)
notify_team(f"Emergency rollback: {prompt_name} -> {stable}")