openrouter-skill

Compare original and translation side by side

🇺🇸

Original

English

🇨🇳

Translation

Chinese

<objective> Enable intelligent routing to Chinese/open-source LLMs through OpenRouter's unified API. Provides model selection guidance, cost optimization, and production patterns for LangChain/LangGraph integration. </objective>

<quick_start>

<objective> 通过OpenRouter的统一API实现对中文/开源大模型的智能路由。为LangChain/LangGraph集成提供模型选择指导、成本优化方案及生产级实践模式。 </objective>

<quick_start>

1. Basic LangChain Setup

1. 基础LangChain配置

python

from langchain_openai import ChatOpenAI
import os

python

from langchain_openai import ChatOpenAI
import os

Any OpenRouter model works with ChatOpenAI

llm = ChatOpenAI( model="deepseek/deepseek-chat", openai_api_key=os.getenv("OPENROUTER_API_KEY"), openai_api_base="https://openrouter.ai/api/v1", default_headers={ "HTTP-Referer": "https://your-app.com", # Optional but recommended "X-Title": "Your App Name" } )

response = llm.invoke("Explain quantum computing in simple terms")

undefined

response = llm.invoke("Explain quantum computing in simple terms")

undefined

2. Vision Analysis (Charts, Documents)

2. 视觉分析（图表、文档）

python

from langchain_core.messages import HumanMessage
import base64

llm = ChatOpenAI(
    model="qwen/qwen-2-vl-72b-instruct",
    openai_api_key=os.getenv("OPENROUTER_API_KEY"),
    openai_api_base="https://openrouter.ai/api/v1"
)

python

from langchain_core.messages import HumanMessage
import base64

llm = ChatOpenAI(
    model="qwen/qwen-2-vl-72b-instruct",
    openai_api_key=os.getenv("OPENROUTER_API_KEY"),
    openai_api_base="https://openrouter.ai/api/v1"
)

From URL

response = llm.invoke([ HumanMessage(content=[ {"type": "text", "text": "Analyze this chart and identify key trends"}, {"type": "image_url", "image_url": {"url": "https://example.com/chart.png"}} ]) ])

From base64

with open("chart.png", "rb") as f: image_data = base64.b64encode(f.read()).decode()

response = llm.invoke([ HumanMessage(content=[ {"type": "text", "text": "What does this chart show?"}, {"type": "image_url", "image_url": {"url": f"data:image/png;base64,{image_data}"}} ]) ])

undefined

with open("chart.png", "rb") as f: image_data = base64.b64encode(f.read()).decode()

response = llm.invoke([ HumanMessage(content=[ {"type": "text", "text": "What does this chart show?"}, {"type": "image_url", "image_url": {"url": f"data:image/png;base64,{image_data}"}} ]) ])

undefined

3. Auto-Routing (Let OpenRouter Choose)

3. 自动路由（由OpenRouter选择）

python

undefined

python

undefined

OpenRouter's Auto model selects the best model for your prompt

llm = ChatOpenAI( model="openrouter/auto", # Powered by NotDiamond openai_api_key=os.getenv("OPENROUTER_API_KEY"), openai_api_base="https://openrouter.ai/api/v1" )


</quick_start>

<success_criteria>
- OpenRouter API key configured and authenticated
- Model selection follows the decision tree (vision -> Qwen-VL, code -> DeepSeek Coder, etc.)
- LangChain ChatOpenAI integration working with correct base URL and headers
- Cost savings of 60-97% vs Western model equivalents for comparable quality
- Fallback chain configured for production reliability
</success_criteria>

<core_concepts>

llm = ChatOpenAI( model="openrouter/auto", # Powered by NotDiamond openai_api_key=os.getenv("OPENROUTER_API_KEY"), openai_api_base="https://openrouter.ai/api/v1" )


</quick_start>

<success_criteria>
- OpenRouter API密钥已配置并验证
- 模型选择遵循决策树（视觉/图表任务→qwen/qwen-2-vl-72b-instruct，代码生成→deepseek/deepseek-coder等）
- LangChain ChatOpenAI与正确的基础URL及请求头集成正常
- 与同质量的西方大模型相比，成本降低60-97%
- 已配置降级链以保障生产环境可靠性
</success_criteria>

<core_concepts>

Model Selection Decision Tree

模型选择决策树

Task Type
│
├─ Vision/Charts ──────────> qwen/qwen-2-vl-72b-instruct ($0.40/M)
├─ Code Generation ────────> deepseek/deepseek-coder ($0.14/$0.28)
├─ Deep Reasoning ─────────> qwen/qwq-32b ($0.15/$0.40)
├─ Long Documents ─────────> moonshot/moonshot-v1-128k ($0.55/M)
├─ Fast/Cheap Tasks ───────> qwen/qwen-2.5-7b-instruct ($0.09/M)
├─ General Analysis ───────> deepseek/deepseek-chat ($0.27/$1.10)
└─ Unknown/Auto ───────────> openrouter/auto

任务类型
│
├─ 视觉/图表 ──────────> qwen/qwen-2-vl-72b-instruct ($0.40/M)
├─ 代码生成 ────────> deepseek/deepseek-coder ($0.14/$0.28)
├─ 深度推理 ─────────> qwen/qwq-32b ($0.15/$0.40)
├─ 长文档处理 ─────────> moonshot/moonshot-v1-128k ($0.55/M)
├─ 快速/低成本任务 ───────> qwen/qwen-2.5-7b-instruct ($0.09/M)
├─ 通用分析 ───────> deepseek/deepseek-chat ($0.27/$1.10)
└─ 未知/自动选择 ───────────> openrouter/auto

Top Chinese LLMs

主流中文大模型

Model	Best For	Cost ($/1M tokens)
`deepseek/deepseek-chat`	General reasoning, analysis	$0.27 in / $1.10 out
`deepseek/deepseek-coder`	Code generation	$0.14 / $0.28
`qwen/qwen-2-vl-72b-instruct`	Vision, charts	$0.40 / $0.40
`qwen/qwen-2.5-7b-instruct`	Fast, cheap tasks	$0.09 / $0.09
`qwen/qwq-32b`	Deep reasoning	$0.15 / $0.40
`moonshot/moonshot-v1-128k`	Long context (128K)	$0.55 / $0.55

模型	适用场景	成本（美元/百万token）
`deepseek/deepseek-chat`	通用推理、分析	$0.27 输入 / $1.10 输出
`deepseek/deepseek-coder`	代码生成	$0.14 / $0.28
`qwen/qwen-2-vl-72b-instruct`	视觉、图表处理	$0.40 / $0.40
`qwen/qwen-2.5-7b-instruct`	快速、低成本任务	$0.09 / $0.09
`qwen/qwq-32b`	深度推理	$0.15 / $0.40
`moonshot/moonshot-v1-128k`	长上下文处理（128K）	$0.55 / $0.55

Cost Comparison vs Western Models

与西方大模型的成本对比

Task	Western Model	Cost	Chinese Model	Cost	Savings
Chat	GPT-4o	$5.00/$15.00	DeepSeek Chat	$0.27/$1.10	95%
Code	Claude Sonnet	$3.00/$15.00	DeepSeek Coder	$0.14/$0.28	95%
Vision	GPT-4o	$5.00/$15.00	Qwen-VL	$0.40/$0.40	97%
Fast	GPT-4o-mini	$0.15/$0.60	Qwen-7B	$0.09/$0.09	60%

任务类型	西方大模型	成本	中文大模型	成本	成本节约率
聊天对话	GPT-4o	$5.00/$15.00	DeepSeek Chat	$0.27/$1.10	95%
代码生成	Claude Sonnet	$3.00/$15.00	DeepSeek Coder	$0.14/$0.28	95%
视觉处理	GPT-4o	$5.00/$15.00	Qwen-VL	$0.40/$0.40	97%
快速任务	GPT-4o-mini	$0.15/$0.60	Qwen-7B	$0.09/$0.09	60%

LangGraph Multi-Model Factory

LangGraph多模型工厂

python

from enum import Enum
from langchain_openai import ChatOpenAI
import os

class ChineseModel(str, Enum):
    DEEPSEEK_CHAT = "deepseek/deepseek-chat"
    DEEPSEEK_CODER = "deepseek/deepseek-coder"
    QWEN_VL = "qwen/qwen-2-vl-72b-instruct"
    QWEN_FAST = "qwen/qwen-2.5-7b-instruct"
    QWQ_REASONING = "qwen/qwq-32b"
    MOONSHOT_LONG = "moonshot/moonshot-v1-128k"
    AUTO = "openrouter/auto"

def create_llm(model: ChineseModel, **kwargs) -> ChatOpenAI:
    """Factory for OpenRouter LLMs with sensible defaults."""
    return ChatOpenAI(
        model=model.value,
        openai_api_key=os.getenv("OPENROUTER_API_KEY"),
        openai_api_base="https://openrouter.ai/api/v1",
        default_headers={
            "HTTP-Referer": os.getenv("APP_URL", "http://localhost"),
            "X-Title": os.getenv("APP_NAME", "LangChain App")
        },
        **kwargs
    )

python

from enum import Enum
from langchain_openai import ChatOpenAI
import os

class ChineseModel(str, Enum):
    DEEPSEEK_CHAT = "deepseek/deepseek-chat"
    DEEPSEEK_CODER = "deepseek/deepseek-coder"
    QWEN_VL = "qwen/qwen-2-vl-72b-instruct"
    QWEN_FAST = "qwen/qwen-2.5-7b-instruct"
    QWQ_REASONING = "qwen/qwq-32b"
    MOONSHOT_LONG = "moonshot/moonshot-v1-128k"
    AUTO = "openrouter/auto"

def create_llm(model: ChineseModel, **kwargs) -> ChatOpenAI:
    """基于合理默认值创建OpenRouter大模型的工厂函数。"""
    return ChatOpenAI(
        model=model.value,
        openai_api_key=os.getenv("OPENROUTER_API_KEY"),
        openai_api_base="https://openrouter.ai/api/v1",
        default_headers={
            "HTTP-Referer": os.getenv("APP_URL", "http://localhost"),
            "X-Title": os.getenv("APP_NAME", "LangChain App")
        },
        **kwargs
    )

Usage

chat_llm = create_llm(ChineseModel.DEEPSEEK_CHAT) vision_llm = create_llm(ChineseModel.QWEN_VL) fast_llm = create_llm(ChineseModel.QWEN_FAST, temperature=0)

undefined

chat_llm = create_llm(ChineseModel.DEEPSEEK_CHAT) vision_llm = create_llm(ChineseModel.QWEN_VL) fast_llm = create_llm(ChineseModel.QWEN_FAST, temperature=0)

undefined

Environment Setup

环境配置

bash

undefined

bash

undefined

.env

OPENROUTER_API_KEY=sk-or-v1-... APP_URL=https://your-app.com # For attribution (optional) APP_NAME=Your App Name # For attribution (optional)


</core_concepts>

<routing>
For detailed information, see:

- `reference/models-catalog.md` - Complete model listing with capabilities
- `reference/routing-strategies.md` - Auto, provider, and custom routing
- `reference/langchain-integration.md` - LangChain/LangGraph patterns
- `reference/cost-optimization.md` - Budget management and caching
- `reference/tool-calling.md` - Function calling patterns
- `reference/multimodal.md` - Vision, PDF, audio support
- `reference/observability.md` - Monitoring and tracing
</routing>

<checklist>
When implementing OpenRouter integration:

- [ ] Set OPENROUTER_API_KEY environment variable
- [ ] Choose appropriate model for task type (see decision tree)
- [ ] Use ChatOpenAI with openai_api_base="https://openrouter.ai/api/v1"
- [ ] Add HTTP-Referer and X-Title headers for attribution
- [ ] Consider cost implications (Chinese models are 10-100x cheaper)
- [ ] Enable streaming for chat interfaces
- [ ] Implement fallback chain for production reliability
- [ ] Set up cost tracking/budget limits
</checklist>

OPENROUTER_API_KEY=sk-or-v1-... APP_URL=https://your-app.com # 用于归因（可选） APP_NAME=Your App Name # 用于归因（可选）


</core_concepts>

<routing>
如需详细信息，请参阅：

- `reference/models-catalog.md` - 包含模型能力的完整模型列表
- `reference/routing-strategies.md` - 自动路由、提供商路由及自定义路由策略
- `reference/langchain-integration.md` - LangChain/LangGraph集成模式
- `reference/cost-optimization.md` - 预算管理与缓存策略
- `reference/tool-calling.md` - 函数调用模式
- `reference/multimodal.md` - 视觉、PDF、音频支持
- `reference/observability.md` - 监控与追踪
</routing>

<checklist>
在集成OpenRouter时，请确认：

- [ ] 已设置OPENROUTER_API_KEY环境变量
- [ ] 根据任务类型选择合适的模型（参考决策树）
- [ ] 使用ChatOpenAI并设置openai_api_base="https://openrouter.ai/api/v1"
- [ ] 添加HTTP-Referer和X-Title请求头以完成归因
- [ ] 考虑成本因素（中文大模型的价格是西方模型的1/10到1/100）
- [ ] 为聊天界面启用流式传输
- [ ] 实现降级链以保障生产环境可靠性
- [ ] 设置成本追踪/预算限制
</checklist>