Agentic RAG 入門

使用 Hugging Face smolagents 的 AI 代理

Adel Nehme

VP of AI Curriculum, DataCamp

Agentic RAG：反覆擷取＋推理

smolagents 中的無狀態 vs. 有狀態工具

函式工具（@tool 裝飾器）

無狀態：呼叫間不保留記憶
呼叫間無法記住你的向量儲存

類別式工具（Tool 子類）

有狀態：可在多次呼叫間保留參照
可儲存複雜物件

類別式工具的結構

from smolagents import Tool

class ToolName(Tool):
    name = "tool_name"

    description = "Clear description for the agent"

    inputs = {
        "parameter_name": {"type": "parameter_type", "description": "Parameter purpose"}
    }
    output_type = "string"

    def __init__(self, custom_parameters):
        super().__init__()
        self.custom_attribute = custom_parameters

    def forward(self, parameter_name):
        # Agent calls this method
        return "processed output"

打造食譜搜尋工具

class RecipeSearchTool(Tool):
    name = "recipe_search" 
    description = "Search cooking documentation for recipes, techniques, and meal planning information"
    inputs = {
        "query": {"type": "string", "description": "Natural language cooking query"}
    }
    output_type = "string"

    def __init__(self, vectorstore, k=6):
        super().__init__()
        self.vectorstore = vectorstore
        self.k = k

    def forward(self, query):
        docs = self.vectorstore.similarity_search(query, k=self.k)
        return "\n\n".join(doc.page_content for doc in docs) or "Nothing found."

料理助理代理：整合實作

# Initialize the retrieval tool
recipe_search = RecipeSearchTool(vector_store)

agent = CodeAgent(
    tools=[recipe_search],
    model=model,
    instructions="Search thoroughly to provide complete recipe answers. 
      If initial results seem incomplete, try different search terms.",
    verbosity_level=1,
    max_steps=8
)

代理執行範例

查詢：如何用專業技巧烹調香草鮭魚？

[步驟 1] 搜尋「salmon herbs cooking techniques」...

[步驟 2] 搜尋「professional salmon preparation」...

[最終答案]

「將魚排擦乾、以香草調味，然後：

平底鍋皮面朝下煎，並以奶油、蒜、百里香澆淋
或以 200°C 與檸檬香草奶油烤 10–12 分鐘」

一起來練習吧！

使用 Hugging Face smolagents 的 AI 代理