1
怎么用LangChain调用ERNIE的function call
AI原生应用开发/技术交流
- 文心大模型
- 插件应用
- LLM
2024.04.081007看过
ERNIE支持functions的模型可以结合LangChain快速开发带自定义插件应用。
千帆平台的官方文档示例是这样的
"functions": [{"name": "get_current_weather","description": "获得指定地点的天气","parameters": {"type": "object","properties": {"location": {"type": "string","description": "省，市名，例如：河北省，石家庄"},"unit": {"type": "string","enum": ["摄氏度", "华氏度"]}},"required": ["location"]}}]
LangChain定义Tool有几种方式，常用的 tool注解和继承BaseTool的类。
使用langchain_core.utils.function_calling包下的方法convert_to_openai_tool，可以把Tool转成描述function的json格式，符合官方的格式要求，很方便。
from langchain_core.tools import toolfrom langchain_core.utils.function_calling import convert_to_openai_toolfrom langchain_community.chat_models.baidu_qianfan_endpoint import QianfanChatEndpoint@tooldef get_current_weather(city: str) -> str:"""获取当前城市的天气"""return f"当前{city}的天气是晴天"tool_json = convert_to_openai_tool(get_current_weather)chat_model = QianfanChatEndpoint(model="ERNIE-3.5-8K", streaming=True).bind(functions=[tool_json["function"]])for chuck in chat_model.stream("北京的天气怎样？"):print(chuck)
这里只是让大模型判断是不是要调用function，如果要调用Tool，还需要增加下一步的逻辑代码。返回如下
content="" additional_kwargs={"finish_reason": "function_call","request_id": "as-4dyf6zy4","object": "chat.completion","search_info": []} response_metadata={"token_usage": {"prompt_tokens": 63,"completion_tokens": 33,"total_tokens": 96},"model_name": "ERNIE-3.5-8K","finish_reason": "function_call","id": "as-4dyf6zy4","object": "chat.completion","created": 1712320443,"result": "","is_truncated": False,"need_clear_history": False,"function_call": {"name": "get_current_weather","thoughts": "用户想要知道北京的天气情况。我可以使用get_current_weather工具来获取北京的天气信息。","arguments": "{"city":"北京"}"},"usage": {"prompt_tokens": 63,"completion_tokens": 33,"total_tokens": 96}}
大模型返回的thoughts挺好的，可惜如果使用LangChain的Message作为输入输出，框架在封装的时候会丢掉，输出里面没办法拿到thoughts。
评论
