|
|
|
|
|
import re
|
|
|
|
|
|
from llm_client import HelloAgentsLLM
|
|
|
|
|
|
from tools import ToolExecutor, search
|
|
|
|
|
|
|
|
|
|
|
|
# (此处省略 REACT_PROMPT_TEMPLATE 的定义)
|
|
|
|
|
|
REACT_PROMPT_TEMPLATE = """
|
|
|
|
|
|
请注意，你是一个有能力调用外部工具的智能助手。
|
|
|
|
|
|
|
|
|
|
|
|
可用工具如下：
|
|
|
|
|
|
{tools}
|
|
|
|
|
|
|
|
|
|
|
|
请严格按照以下格式进行回应：
|
|
|
|
|
|
|
|
|
|
|
|
Thought: 你的思考过程，用于分析问题、拆解任务和规划下一步行动。
|
|
|
|
|
|
Action: 你决定采取的行动，必须是以下格式之一：
|
|
|
|
|
|
- `{{tool_name}}[{{tool_input}}]`：调用一个可用工具。
|
|
|
|
|
|
- `Finish[最终答案]`：当你认为已经获得最终答案时。
|
|
|
|
|
|
|
|
|
|
|
|
现在，请开始解决以下问题：
|
|
|
|
|
|
Question: {question}
|
|
|
|
|
|
History: {history}
|
|
|
|
|
|
"""
|
|
|
|
|
|
|
|
|
|
|
|
class ReActAgent:
|
|
|
|
|
|
def __init__(self, llm_client: HelloAgentsLLM, tool_executor: ToolExecutor, max_steps: int = 5):
|
|
|
|
|
|
self.llm_client = llm_client
|
|
|
|
|
|
self.tool_executor = tool_executor
|
|
|
|
|
|
self.max_steps = max_steps
|
|
|
|
|
|
self.history = []
|
|
|
|
|
|
|
|
|
|
|
|
def run(self, question: str):
|
|
|
|
|
|
self.history = []
|
|
|
|
|
|
current_step = 0
|
|
|
|
|
|
|
|
|
|
|
|
while current_step < self.max_steps:
|
|
|
|
|
|
current_step += 1
|
|
|
|
|
|
print(f"\n--- 第 {current_step} 步 ---")
|
|
|
|
|
|
|
|
|
|
|
|
tools_desc = self.tool_executor.getAvailableTools()
|
|
|
|
|
|
history_str = "\n".join(self.history)
|
|
|
|
|
|
prompt = REACT_PROMPT_TEMPLATE.format(tools=tools_desc, question=question, history=history_str)
|
|
|
|
|
|
|
|
|
|
|
|
messages = [{"role": "user", "content": prompt}]
|
|
|
|
|
|
response_text = self.llm_client.think(messages=messages)
|
|
|
|
|
|
if not response_text:
|
|
|
|
|
|
print("错误：LLM未能返回有效响应。"); break
|
|
|
|
|
|
|
|
|
|
|
|
thought, action = self._parse_output(response_text)
|
|
|
|
|
|
if thought: print(f"🤔 思考: {thought}")
|
|
|
|
|
|
if not action: print("警告：未能解析出有效的Action，流程终止。"); break
|
|
|
|
|
|
|
|
|
|
|
|
if action.startswith("Finish"):
|
|
|
|
|
|
final_answer = self._parse_action_input(action)
|
|
|
|
|
|
print(f"🎉 最终答案: {final_answer}")
|
|
|
|
|
|
return final_answer
|
|
|
|
|
|
|
|
|
|
|
|
tool_name, tool_input = self._parse_action(action)
|
|
|
|
|
|
if not tool_name or not tool_input:
|
|
|
|
|
|
self.history.append("Observation: 无效的Action格式，请检查。"); continue
|
|
|
|
|
|
|
|
|
|
|
|
print(f"🎬 行动: {tool_name}[{tool_input}]")
|
|
|
|
|
|
tool_function = self.tool_executor.getTool(tool_name)
|
|
|
|
|
|
observation = tool_function(tool_input) if tool_function else f"错误：未找到名为 '{tool_name}' 的工具。"
|
|
|
|
|
|
|
|
|
|
|
|
print(f"👀 观察: {observation}")
|
|
|
|
|
|
self.history.append(f"Action: {action}")
|
|
|
|
|
|
self.history.append(f"Observation: {observation}")
|
|
|
|
|
|
|
|
|
|
|
|
print("已达到最大步数，流程终止。")
|
|
|
|
|
|
return None
|
|
|
|
|
|
|
|
|
|
|
|
def _parse_output(self, text: str):
|
|
|
|
|
|
thought_match = re.search(r"Thought: (.*)", text)
|
|
|
|
|
|
action_match = re.search(r"Action: (.*)", text)
|
|
|
|
|
|
thought = thought_match.group(1).strip() if thought_match else None
|
|
|
|
|
|
action = action_match.group(1).strip() if action_match else None
|
|
|
|
|
|
return thought, action
|
|
|
|
|
|
|
|
|
|
|
|
def _parse_action(self, action_text: str):
|
|
|
|
|
|
match = re.match(r"(\w+)\[(.*)\]", action_text)
|
|
|
|
|
|
return (match.group(1), match.group(2)) if match else (None, None)
|
|
|
|
|
|
|
|
|
|
|
|
def _parse_action_input(self, action_text: str):
|
|
|
|
|
|
match = re.match(r"\w+\[(.*)\]", action_text)
|
|
|
|
|
|
return match.group(1) if match else ""
|
|
|
|
|
|
|
|
|
|
|
|
if __name__ == '__main__':
|
|
|
|
|
|
llm = HelloAgentsLLM()
|
|
|
|
|
|
tool_executor = ToolExecutor()
|
|
|
|
|
|
search_desc = "一个网页搜索引擎。当你需要回答关于时事、事实以及在你的知识库中找不到的信息时，应使用此工具。"
|
|
|
|
|
|
tool_executor.registerTool("Search", search_desc, search)
|
|
|
|
|
|
agent = ReActAgent(llm_client=llm, tool_executor=tool_executor)
|
|
|
|
|
|
question = "华为最新的手机是哪一款？它的主要卖点是什么？"
|
|
|
|
|
|
agent.run(question)
|
