202 lines
6.9 KiB
Python
202 lines
6.9 KiB
Python
"""
|
||
ReAct Agent（基于 DeepSeek Function Calling 实现）
|
||
|
||
ReAct = Reasoning + Acting：模型先"想"（决定是否调用工具、调用哪个），
|
||
再"做"（执行工具），拿到"观察结果"后继续下一轮思考，直到给出最终答案。
|
||
"""
|
||
|
||
import os
|
||
import json
|
||
from openai import OpenAI
|
||
from dotenv import load_dotenv
|
||
|
||
load_dotenv()
|
||
|
||
client = OpenAI(
|
||
api_key=os.environ.get("DEEPSEEK_API_KEY"),
|
||
base_url="https://api.deepseek.com",
|
||
)
|
||
|
||
MODEL = "deepseek-chat" # 换成你实际可用的模型名
|
||
MAX_STEPS = 8 # 防止死循环的最大推理步数
|
||
|
||
|
||
# ============================================================
|
||
# 1. 工具实现层：每个工具就是一个普通的 Python 函数
|
||
# ============================================================
|
||
def get_weather(location: str) -> str:
|
||
"""模拟天气查询接口"""
|
||
fake_db = {
|
||
"杭州": "30℃，晴，东南风 2 级",
|
||
"北京": "25℃，多云",
|
||
"上海": "28℃，小雨",
|
||
}
|
||
return fake_db.get(location, f"{location}：暂无该城市天气数据")
|
||
|
||
|
||
def calculator(expression: str) -> str:
|
||
"""一个安全的四则运算计算器"""
|
||
allowed = set("0123456789+-*/(). ")
|
||
if not set(expression) <= allowed:
|
||
return "错误：表达式包含非法字符"
|
||
try:
|
||
return str(eval(expression, {"__builtins__": {}}, {}))
|
||
except Exception as e:
|
||
return f"计算失败：{e}"
|
||
|
||
|
||
# 工具名 -> 函数 的注册表，Agent 靠它做分发
|
||
TOOL_REGISTRY = {
|
||
"get_weather": get_weather,
|
||
"calculator": calculator,
|
||
}
|
||
|
||
|
||
# ============================================================
|
||
# 2. 工具描述层：告诉模型有哪些工具可用（JSON Schema）
|
||
# ============================================================
|
||
TOOLS_SCHEMA = [
|
||
{
|
||
"type": "function",
|
||
"function": {
|
||
"name": "get_weather",
|
||
"description": "获取指定城市的当前天气",
|
||
"parameters": {
|
||
"type": "object",
|
||
"properties": {
|
||
"location": {"type": "string", "description": "城市名，如杭州"},
|
||
},
|
||
"required": ["location"],
|
||
},
|
||
},
|
||
},
|
||
{
|
||
"type": "function",
|
||
"function": {
|
||
"name": "calculator",
|
||
"description": "计算一个数学表达式，例如 '30 - 25'",
|
||
"parameters": {
|
||
"type": "object",
|
||
"properties": {
|
||
"expression": {"type": "string", "description": "要计算的数学表达式"},
|
||
},
|
||
"required": ["expression"],
|
||
},
|
||
},
|
||
},
|
||
]
|
||
|
||
|
||
# ============================================================
|
||
# 3. 系统提示词：定义 Agent 的行为方式
|
||
# ============================================================
|
||
SYSTEM_PROMPT = """你是一个可以使用工具的智能助手。
|
||
|
||
工作要求：
|
||
1. 先思考问题需要哪些信息，判断是否需要调用工具。
|
||
2. 需要外部信息（天气、计算等）时，必须调用工具，不要凭空编造。
|
||
3. 可以连续多次调用工具；每一步只做当前最必要的事。
|
||
4. 拿到工具结果后，用简洁的中文给出最终答案。
|
||
"""
|
||
|
||
|
||
# ============================================================
|
||
# 4. 核心：ReAct 循环
|
||
# ============================================================
|
||
def run_react_agent(user_query: str, max_steps: int = MAX_STEPS, verbose: bool = True):
|
||
"""
|
||
执行 ReAct 循环：
|
||
思考(Thought) -> 行动(Action) -> 观察(Observation) -> 再思考 ... -> 最终答案
|
||
返回 (最终答案, 完整消息历史)
|
||
"""
|
||
messages = [
|
||
{"role": "system", "content": SYSTEM_PROMPT},
|
||
{"role": "user", "content": user_query},
|
||
]
|
||
|
||
for step in range(1, max_steps + 1):
|
||
if verbose:
|
||
print(f"\n{'=' * 20} Step {step} {'=' * 20}")
|
||
|
||
# ---------- Thought：让模型决定下一步做什么 ----------
|
||
response = client.chat.completions.create(
|
||
model=MODEL,
|
||
messages=messages,
|
||
tools=TOOLS_SCHEMA,
|
||
tool_choice="auto",
|
||
)
|
||
assistant_msg = response.choices[0].message
|
||
|
||
# 把模型的回复写入历史（必须手动构造，content 为 None 时补成 ""）
|
||
msg_dict = {"role": "assistant", "content": assistant_msg.content or ""}
|
||
if assistant_msg.tool_calls:
|
||
msg_dict["tool_calls"] = [tc.model_dump() for tc in assistant_msg.tool_calls]
|
||
messages.append(msg_dict)
|
||
|
||
if verbose and assistant_msg.content:
|
||
print(f"[Thought] {assistant_msg.content}")
|
||
|
||
# ---------- 终止条件：模型不再请求工具，直接给出答案 ----------
|
||
if not assistant_msg.tool_calls:
|
||
final_answer = assistant_msg.content or ""
|
||
if verbose:
|
||
print(f"[Final] {final_answer}")
|
||
return final_answer, messages
|
||
|
||
# ---------- Action + Observation：执行所有工具调用 ----------
|
||
for tool_call in assistant_msg.tool_calls:
|
||
name = tool_call.function.name
|
||
|
||
# 注意：arguments 是 JSON 字符串，必须解析成 dict 再传参
|
||
try:
|
||
args = json.loads(tool_call.function.arguments or "{}")
|
||
except json.JSONDecodeError:
|
||
args = {}
|
||
|
||
if verbose:
|
||
print(f"[Action] {name}({json.dumps(args, ensure_ascii=False)})")
|
||
|
||
func = TOOL_REGISTRY.get(name)
|
||
if func is None:
|
||
result = f"错误：未知工具 {name}"
|
||
else:
|
||
try:
|
||
result = func(**args)
|
||
except Exception as e:
|
||
result = f"工具执行异常：{e}"
|
||
|
||
result = str(result)
|
||
if verbose:
|
||
print(f"[Observation] {result}")
|
||
|
||
# 工具结果必须带上 tool_call_id，一一对应
|
||
messages.append({
|
||
"role": "tool",
|
||
"tool_call_id": tool_call.id,
|
||
"content": result,
|
||
})
|
||
|
||
# 超过最大步数仍未收敛
|
||
fallback = "抱歉，我在限定步骤内没能得出结论。"
|
||
if verbose:
|
||
print(f"[Final] {fallback}")
|
||
return fallback, messages
|
||
|
||
|
||
# ============================================================
|
||
# 5. 运行入口
|
||
# ============================================================
|
||
if __name__ == "__main__":
|
||
# 单轮测试
|
||
for q in ["杭州天气怎么样？", "杭州和北京温差多少度？"]:
|
||
print(f"\n\n########## 用户：{q} ##########")
|
||
answer, history = run_react_agent(q)
|
||
print(f"回答：{answer}")
|
||
|
||
# 如果想做成多轮对话的交互式 Agent，可以这样：
|
||
# while True:
|
||
# q = input("\n你：")
|
||
# if q in ("exit", "quit"):
|
||
# break
|
||
# answer, _ = run_react_agent(q)
|
||
# print("助手：", answer) |