|
|
@@ -1276,6 +1276,72 @@ print(f"数学专用Agent结果: {math_result}")
|
|
|
<img src="https://raw.githubusercontent.com/datawhalechina/Hello-Agents/main/docs/images/7-figures/table-02.png" alt="" width="90%"/>
|
|
|
</div>
|
|
|
|
|
|
+### 7.4.5 FunctionCallAgent
|
|
|
+
|
|
|
+FunctionCallAgent是hello-agents在0.2.8之后引入的Agent,它基于OpenAI原生函数调用机制的Agent,展示了如何使用OpenAI的函数调用机制来构建Agent。
|
|
|
+它支持以下功能:
|
|
|
+_build_tool_schemas:通过工具的description构建OpenAI的function calling schema
|
|
|
+_extract_message_content:从OpenAI的响应中提取文本
|
|
|
+_parse_function_call_arguments:解析模型返回的JSON字符串参数
|
|
|
+_convert_parameter_types:转换参数类型
|
|
|
+
|
|
|
+这些功能可以使其具备原生的OpenAI Functioncall的能力,对比使用prompt约束的方式,具备更强的鲁棒性。
|
|
|
+```python
|
|
|
+def _invoke_with_tools(self, messages: list[dict[str, Any]], tools: list[dict[str, Any]], tool_choice: Union[str, dict], **kwargs):
|
|
|
+ """调用底层OpenAI客户端执行函数调用"""
|
|
|
+ client = getattr(self.llm, "_client", None)
|
|
|
+ if client is None:
|
|
|
+ raise RuntimeError("HelloAgentsLLM 未正确初始化客户端,无法执行函数调用。")
|
|
|
+
|
|
|
+ client_kwargs = dict(kwargs)
|
|
|
+ client_kwargs.setdefault("temperature", self.llm.temperature)
|
|
|
+ if self.llm.max_tokens is not None:
|
|
|
+ client_kwargs.setdefault("max_tokens", self.llm.max_tokens)
|
|
|
+
|
|
|
+ return client.chat.completions.create(
|
|
|
+ model=self.llm.model,
|
|
|
+ messages=messages,
|
|
|
+ tools=tools,
|
|
|
+ tool_choice=tool_choice,
|
|
|
+ **client_kwargs,
|
|
|
+ )
|
|
|
+
|
|
|
+#内部逻辑是对Openai 原生的functioncall作再封装
|
|
|
+#OpenAI 原生functioncall示例
|
|
|
+from openai import OpenAI
|
|
|
+client = OpenAI()
|
|
|
+
|
|
|
+tools = [
|
|
|
+ {
|
|
|
+ "type": "function",
|
|
|
+ "function": {
|
|
|
+ "name": "get_current_weather",
|
|
|
+ "description": "Get the current weather in a given location",
|
|
|
+ "parameters": {
|
|
|
+ "type": "object",
|
|
|
+ "properties": {
|
|
|
+ "location": {
|
|
|
+ "type": "string",
|
|
|
+ "description": "The city and state, e.g. San Francisco, CA",
|
|
|
+ },
|
|
|
+ "unit": {"type": "string", "enum": ["celsius", "fahrenheit"]},
|
|
|
+ },
|
|
|
+ "required": ["location"],
|
|
|
+ },
|
|
|
+ }
|
|
|
+ }
|
|
|
+]
|
|
|
+messages = [{"role": "user", "content": "What's the weather like in Boston today?"}]
|
|
|
+completion = client.chat.completions.create(
|
|
|
+ model="gpt-5",
|
|
|
+ messages=messages,
|
|
|
+ tools=tools,
|
|
|
+ tool_choice="auto"
|
|
|
+)
|
|
|
+
|
|
|
+print(completion)
|
|
|
+```
|
|
|
+
|
|
|
## 7.5 工具系统
|
|
|
|
|
|
本节内容将在前面构建的Agent基础架构上,深入探讨工具系统的设计与实现。我们将从基础设施建设开始,逐步深入到自定义开发设计。本节的学习目标围绕以下三个核心方面展开:
|
|
|
@@ -1392,6 +1458,57 @@ def get_tools_description(self) -> str:
|
|
|
````
|
|
|
这个方法生成的描述字符串可以直接用于构建Agent的提示词,让Agent了解可用的工具。
|
|
|
|
|
|
+````python
|
|
|
+def to_openai_schema(self) -> Dict[str, Any]:
|
|
|
+ """转换为 OpenAI function calling schema 格式
|
|
|
+
|
|
|
+ 用于 FunctionCallAgent,使工具能够被 OpenAI 原生 function calling 使用
|
|
|
+
|
|
|
+ Returns:
|
|
|
+ 符合 OpenAI function calling 标准的 schema
|
|
|
+ """
|
|
|
+ parameters = self.get_parameters()
|
|
|
+
|
|
|
+ # 构建 properties
|
|
|
+ properties = {}
|
|
|
+ required = []
|
|
|
+
|
|
|
+ for param in parameters:
|
|
|
+ # 基础属性定义
|
|
|
+ prop = {
|
|
|
+ "type": param.type,
|
|
|
+ "description": param.description
|
|
|
+ }
|
|
|
+
|
|
|
+ # 如果有默认值,添加到描述中(OpenAI schema 不支持 default 字段)
|
|
|
+ if param.default is not None:
|
|
|
+ prop["description"] = f"{param.description} (默认: {param.default})"
|
|
|
+
|
|
|
+ # 如果是数组类型,添加 items 定义
|
|
|
+ if param.type == "array":
|
|
|
+ prop["items"] = {"type": "string"} # 默认字符串数组
|
|
|
+
|
|
|
+ properties[param.name] = prop
|
|
|
+
|
|
|
+ # 收集必需参数
|
|
|
+ if param.required:
|
|
|
+ required.append(param.name)
|
|
|
+
|
|
|
+ return {
|
|
|
+ "type": "function",
|
|
|
+ "function": {
|
|
|
+ "name": self.name,
|
|
|
+ "description": self.description,
|
|
|
+ "parameters": {
|
|
|
+ "type": "object",
|
|
|
+ "properties": properties,
|
|
|
+ "required": required
|
|
|
+ }
|
|
|
+ }
|
|
|
+ }
|
|
|
+````
|
|
|
+这个方法生成的schema可以直接用于原生的OpenAI SDK的工具调用。
|
|
|
+
|
|
|
### 7.5.2 自定义工具开发
|
|
|
|
|
|
有了基础设施后,我们来看看如何开发一个完整的自定义工具。数学计算工具是一个很好的例子,因为它简单直观,最直接的方式是使用ToolRegistry的函数注册功能。
|