import anthropic

client = anthropic.Anthropic(
    api_key='your-derouter-api-key',
    base_url='https://api.derouter.ai/proxy/v1'
)

def smart_route(task_type, content):
    '''根据任务类型自动选择最经济的模型'''
    model_map = {
        'classify': 'claude-haiku-4-5-20251001',   # 分类用 Haiku
        'generate': 'claude-sonnet-4-6',             # 生成用 Sonnet
        'reason': 'claude-opus-4-6',                 # 推理用 Opus
    }

    model = model_map.get(task_type, 'claude-sonnet-4-6')

    msg = client.messages.create(
        model=model,
        max_tokens=2048,
        messages=[{'role': 'user', 'content': content}]
    )
    return msg.content[0].text

# 简单分类 → Haiku（最便宜）
category = smart_route('classify', '判断这条评论的情感：这个产品太棒了！')

# 内容生成 → Sonnet（性价比最高）
article = smart_route('generate', '写一篇关于远程办公的文章')

# 复杂推理 → Opus（最强）
analysis = smart_route('reason', '分析这份财报数据...')

# ❌ 冗长的 System Prompt（1200 tokens）
bad_prompt = '''
You are a helpful AI assistant working for our company.
Our company is called TechCorp and we specialize in providing
software solutions for small and medium businesses. You should
always be polite, professional, and helpful. When answering
questions, try to be concise but thorough. If you don't know
something, admit it rather than making things up. You should
respond in the same language the user writes in...
（还有更多...）
'''

# ✅ 精简的 System Prompt（200 tokens）
good_prompt = '''TechCorp AI assistant. Rules:
- Concise, professional
- Match user language
- Cite sources when possible
- Say \

def truncate_history(messages, max_turns=10):
    '''保留最近 N 轮对话，旧对话用摘要替代'''
    if len(messages) <= max_turns * 2:
        return messages

    # 旧对话让 AI 生成摘要
    old_messages = messages[:-max_turns * 2]
    summary = summarize(old_messages)  # 用 Haiku 生成摘要

    return [
        {'role': 'user', 'content': f'对话摘要：{summary}'},
        {'role': 'assistant', 'content': '好的，我了解之前的对话内容。'},
        *messages[-max_turns * 2:]
    ]

import hashlib
import json
import redis

r = redis.Redis()

def cached_completion(prompt, model='claude-sonnet-4-6', ttl=3600):
    '''结果缓存：相同输入直接返回缓存结果'''
    cache_key = f'llm:{hashlib.md5(f\