llm_adaptive_router

LLM自適應路由器

LLM自適應路由器是一個Python軟件包，可以基於查詢內容啟用動態模型選擇。它使用有效的矢量搜索進行初始分類和基於LLM的細粒度選擇，以進行複雜情況。路由器可以從反饋中適應和學習，使其適用於廣泛的應用。

特徵

• 基於查詢內容的動態模型選擇
• 有效的矢量搜索初始分類
• 基於LLM的複雜情況的細粒度選擇
• 從反饋中自適應學習
• 路線和模型的靈活配置
• 與Langchain和各種LLM提供商易於集成

安裝

您可以使用PIP安裝LLM自適應路由器：

pip3 install llm-adaptive-router

快速開始

這是如何使用LLM自適應路由器的一個基本示例：

 from llm_adaptive_router import AdaptiveRouter , RouteMetadata
from langchain_chroma import Chroma
from langchain_openai import OpenAIEmbeddings , ChatOpenAI
from dotenv import load_dotenv
load_dotenv ()

gpt_3_5_turbo = ChatOpenAI ( model = "gpt-3.5-turbo" )
mini = ChatOpenAI ( model = "gpt-4o-mini" )
gpt_4 = ChatOpenAI ( model = "gpt-4" )

routes = {
    "general" : RouteMetadata (
        invoker = gpt_3_5_turbo ,
        capabilities = [ "general knowledge" ],
        cost = 0.002 ,
        example_sentences = [ "What is the capital of France?" , "Explain photosynthesis." ]
    ),
    "mini" : RouteMetadata (
        invoker = mini ,
        capabilities = [ "general knowledge" ],
        cost = 0.002 ,
        example_sentences = [ "What is the capital of France?" , "Explain photosynthesis." ]
        
    ),
    "math" : RouteMetadata (
        invoker = gpt_4 ,
        capabilities = [ "advanced math" , "problem solving" ],
        cost = 0.01 ,
        example_sentences = [ "Solve this differential equation." , "Prove the Pythagorean theorem." ]
    )
}

llm = ChatOpenAI ( model = "gpt-3.5-turbo" )

router = AdaptiveRouter (
    vectorstore = Chroma ( embedding_function = OpenAIEmbeddings ()),
    llm = llm ,
    embeddings = OpenAIEmbeddings (),
    routes = routes
)

query = "How are you"
query2 = "Write a Python function to hello world"
selected_model_route = router . route ( query )
selected_model_name = selected_model_route
print ( selected_model_name )
invoker = selected_model_route . invoker
response = invoker . invoke ( query )

print ( f"Response: { response } " )

詳細的用法

創建路線元數據

使用create_route_metadata函數來定義路由：

 from llm_adaptive_router import create_route_metadata

route = create_route_metadata (
    invoker = model_function ,
    capabilities = [ "capability1" , "capability2" ],
    cost = 0.01 ,
    example_sentences = [ "Example query 1" , "Example query 2" ],
    additional_info = { "key" : "value" }
)

初始化Adaptiverouter

使用您的配置路由創建一個AdaptiveRouter的實例：

 router = AdaptiveRouter (
    vectorstore = your_vectorstore ,
    llm = your_llm ,
    embeddings = your_embeddings ,
    routes = your_routes
)

路由查詢

使用route方法為查詢選擇適當的模型：

 selected_model_route = router . route ( "Your query here" )
selected_model_name = selected_model_route . model
invoker = selected_model_route . invoker
response = invoker . invoke ( "Your query here" )

添加反饋

通過提供反饋來提高路由器的性能：

 router . add_feedback ( query , selected_model , performance_score )

高級功能

• 自定義矢量商店：LLM自適應路由器支持各種向量商店。您可以使用任何實現Langchain實現VectorStore店接口的矢量存儲。
• 動態路由更新：您可以動態添加或刪除路由：

 router . add_route ( "new_route" , new_route_metadata )
router . remove_route ( "old_route" )

• 調整路由器行為：微調路由器的行為：

 router . set_complexity_threshold ( 0.8 )
router . set_update_frequency ( 200 )