月省千元的六层AI省钱架构实测

flowchart TD
    A[用户请求] --> B{第1层: 语义缓存<br/>命中?}
    B -->|命中| C[直接返回缓存<br/>成本: $0]
    B -->|未命中| D{第2层: 模型路由器<br/>复杂度判断}
    D -->|简单任务| E[小模型<br/>Haiku/GPT-4o-mini<br/>成本: $0.15-0.8/MTok]
    D -->|中等任务| F[中档模型<br/>Sonnet/GPT-4o<br/>成本: $2.5-3/MTok]
    D -->|复杂任务| G[大模型<br/>Opus/o1<br/>成本: $15/MTok]
    E --> H{第3层: 批处理<br/>实时?}
    F --> H
    H -->|非实时| I[Batch API<br/>省50%]
    H -->|实时| J{第4层: 提示词缓存<br/>命中?}
    J -->|命中| K[输入省90%]
    J -->|未命中| L{第5层: 降价时段<br/>可用?}
    I --> M[第6层: 开源自托管兜底]
    L --> M
    G --> J
    K --> N[返回结果+写入缓存]
    M --> N
    style C fill:#4CAF50,color:#fff
    style E fill:#8BC34A,color:#fff
    style F fill:#FFC107,color:#000
    style G fill:#FF5722,color:#fff
    style I fill:#2196F3,color:#fff
    style K fill:#9C27B0,color:#fff
    style M fill:#607D8B,color:#fff

from gptcache import Cache
from gptcache.adapter import openai
from gptcache.embedding import OpenAI as EmbedOpenAI
from gptcache.similarity_evaluation import Cosine

cache = Cache()
cache.init(
    embedding_func=EmbedOpenAI().to_embeddings,
    similarity_evaluation=Cosine(),
    config={"similarity_threshold": 0.85}  # 相似度85%以上命中缓存
)

# 替换原openai调用，其他代码完全不变
response = openai.ChatCompletion.create(
    model="gpt-4o",
    messages=[{"role": "user", "content": user_query}],
    cache=cache  # 就多这一个参数
)

import litellm

# 定义路由策略
router = litellm.Router(
    model_list=[
        {
            "model_name": "cheap",
            "litellm_params": {
                "model": "anthropic/claude-3-5-haiku-20241022",
                "api_key": "sk-ant-xxx"
            },
        },
        {
            "model_name": "medium",
            "litellm_params": {
                "model": "openai/gpt-4o",
                "api_key": "sk-xxx"
            },
        },
        {
            "model_name": "powerful",
            "litellm_params": {
                "model": "anthropic/claude-opus-4-20250514",
                "api_key": "sk-ant-xxx"
            },
        },
    ],
    fallbacks=[{"powerful": ["medium"]}, {"medium": ["cheap"]}],
    context_window_fallbacks=[{"cheap": ["medium"]}]
)

# 在应用层做简单分类
def get_model_for_query(query):
    simple_keywords = ["总结", "翻译", "改写", "格式化", "提取"]
    if any(kw in query for kw in simple_keywords):
        return "cheap"
    complex_keywords = ["分析", "推理", "debug", "架构设计", "比较"]
    if any(kw in query for kw in complex_keywords):
        return "powerful"
    return "medium"

graph LR
    A[你的任务] --> B{需要实时响应?}
    B -->|是| C[走实时API]
    B -->|否| D{吞吐量>1000/天?}
    D -->|否| C
    D -->|是| E{延迟容忍>1h?}
    E -->|否| C
    E -->|是| F[✅ 用Batch API<br/>省50%]

    G[典型Batch场景] --> H[文档分类/标注]
    G --> I[批量翻译]
    G --> J[内容审核]
    G --> K[数据清洗/抽取]
    G --> L[SEO描述生成]

from openai import OpenAI
client = OpenAI()

# 第1步：准备批量请求JSONL文件
import json

requests = []
for i, text in enumerate(documents_to_process):
    requests.append({
        "custom_id": f"doc-{i}",
        "method": "POST",
        "url": "/v1/chat/completions",
        "body": {
            "model": "gpt-4o",
            "messages": [
                {"role": "system", "content": "分类以下文档"},
                {"role": "user", "content": text}
            ]
        }
    })

# 写入JSONL
with open("batch_requests.jsonl", "w") as f:
    for req in requests:
        f.write(json.dumps(req) + "\n")

# 第2步：上传文件
batch_file = client.files.create(
    file=open("batch_requests.jsonl", "rb"),
    purpose="batch"
)

# 第3步：创建批处理任务
batch = client.batches.create(
    input_file_id=batch_file.id,
    endpoint="/v1/chat/completions",
    completion_window="24h"
)

print(f"批处理ID: {batch.id}")
# 检查状态: client.batches.retrieve(batch.id)

import anthropic

client = anthropic.Anthropic()

# 关键：把系统提示词标记为可缓存
response = client.messages.create(
    model="claude-sonnet-4-20250514",
    max_tokens=1024,
    system=[
        {
            "type": "text",
            "text": "你是一个专业的...（很长的系统提示词，比如5000 tokens）",
            "cache_control": {"type": "ephemeral"}  # ← 这一行开启缓存
        }
    ],
    messages=[
        {"role": "user", "content": "用户的具体问题"}
    ]
)

pie title "每月免费额度分布（开发者个人）"
    "Google AI Studio (Gemini)" : 40
    "Groq (每日免费)" : 20
    "Cerebras (每日免费)" : 15
    "DeepSeek (注册赠送)" : 15
    "OpenRouter (免费模型)" : 10

# 一行命令部署
ollama run qwen2.5-coder:14b

# 对比：OpenAI API同样任务月费$20+，本地完全免费

graph TD
    subgraph "月预算 < $50"
        A1[Google AI Studio免费tier]
        A2[DeepSeek API]
        A3[Ollama本地部署]
    end

    subgraph "月预算 $50-200"
        B1[Claude Sonnet + Haiku路由]
        B2[Batch API处理批量任务]
        B3[提示词缓存]
    end

    subgraph "月预算 $200-1000"
        C1[LiteLLM智能路由]
        C2[语义缓存层]
        C3[多供应商分散]
    end

    subgraph "月预算 > $1000"
        D1[全套六层架构]
        D2[考虑自托管关键模型]
        D3[企业级协议谈判]
    end