Spaces:
Sleeping
Sleeping
Upload 2 files
Browse files- agent.py +233 -0
- system_prompt.txt +32 -0
agent.py
ADDED
|
@@ -0,0 +1,233 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
import os
|
| 2 |
+
import time
|
| 3 |
+
import json
|
| 4 |
+
from dotenv import load_dotenv
|
| 5 |
+
from langgraph.graph import StateGraph, END
|
| 6 |
+
from langchain_google_genai import ChatGoogleGenerativeAI
|
| 7 |
+
from langchain_community.tools import DuckDuckGoSearchRun
|
| 8 |
+
from langchain_community.document_loaders import WikipediaLoader, ArxivLoader
|
| 9 |
+
from langchain_core.messages import SystemMessage, AIMessage, HumanMessage
|
| 10 |
+
from langchain_core.tools import tool
|
| 11 |
+
from tenacity import retry, stop_after_attempt, wait_exponential
|
| 12 |
+
from typing import TypedDict, Annotated, Sequence
|
| 13 |
+
import operator
|
| 14 |
+
|
| 15 |
+
# Load environment variables
|
| 16 |
+
load_dotenv()
|
| 17 |
+
google_api_key = os.getenv("GOOGLE_API_KEY") or os.environ.get("GOOGLE_API_KEY")
|
| 18 |
+
if not google_api_key:
|
| 19 |
+
raise ValueError("Missing GOOGLE_API_KEY environment variable")
|
| 20 |
+
|
| 21 |
+
# --- Math Tools ---
|
| 22 |
+
@tool
|
| 23 |
+
def multiply(a: int, b: int) -> int:
|
| 24 |
+
"""Multiply two integers."""
|
| 25 |
+
return a * b
|
| 26 |
+
|
| 27 |
+
@tool
|
| 28 |
+
def add(a: int, b: int) -> int:
|
| 29 |
+
"""Add two integers."""
|
| 30 |
+
return a + b
|
| 31 |
+
|
| 32 |
+
@tool
|
| 33 |
+
def subtract(a: int, b: int) -> int:
|
| 34 |
+
"""Subtract b from a."""
|
| 35 |
+
return a - b
|
| 36 |
+
|
| 37 |
+
@tool
|
| 38 |
+
def divide(a: int, b: int) -> float:
|
| 39 |
+
"""Divide a by b, error on zero."""
|
| 40 |
+
if b == 0:
|
| 41 |
+
raise ValueError("Cannot divide by zero.")
|
| 42 |
+
return a / b
|
| 43 |
+
|
| 44 |
+
@tool
|
| 45 |
+
def modulus(a: int, b: int) -> int:
|
| 46 |
+
"""Compute a mod b."""
|
| 47 |
+
return a % b
|
| 48 |
+
|
| 49 |
+
# --- Browser Tools ---
|
| 50 |
+
@tool
|
| 51 |
+
def wiki_search(query: str) -> str:
|
| 52 |
+
"""Search Wikipedia and return up to 3 relevant documents."""
|
| 53 |
+
try:
|
| 54 |
+
docs = WikipediaLoader(query=query, load_max_docs=3).load()
|
| 55 |
+
if not docs:
|
| 56 |
+
return "No Wikipedia results found."
|
| 57 |
+
|
| 58 |
+
results = []
|
| 59 |
+
for doc in docs:
|
| 60 |
+
title = doc.metadata.get('title', 'Unknown Title')
|
| 61 |
+
content = doc.page_content[:2000] # Limit content length
|
| 62 |
+
results.append(f"Title: {title}\nContent: {content}")
|
| 63 |
+
|
| 64 |
+
return "\n\n---\n\n".join(results)
|
| 65 |
+
except Exception as e:
|
| 66 |
+
return f"Wikipedia search error: {str(e)}"
|
| 67 |
+
|
| 68 |
+
@tool
|
| 69 |
+
def arxiv_search(query: str) -> str:
|
| 70 |
+
"""Search Arxiv and return up to 3 relevant papers."""
|
| 71 |
+
try:
|
| 72 |
+
docs = ArxivLoader(query=query, load_max_docs=3).load()
|
| 73 |
+
if not docs:
|
| 74 |
+
return "No arXiv papers found."
|
| 75 |
+
|
| 76 |
+
results = []
|
| 77 |
+
for doc in docs:
|
| 78 |
+
title = doc.metadata.get('Title', 'Unknown Title')
|
| 79 |
+
authors = ", ".join(doc.metadata.get('Authors', []))
|
| 80 |
+
content = doc.page_content[:2000] # Limit content length
|
| 81 |
+
results.append(f"Title: {title}\nAuthors: {authors}\nContent: {content}")
|
| 82 |
+
|
| 83 |
+
return "\n\n---\n\n".join(results)
|
| 84 |
+
except Exception as e:
|
| 85 |
+
return f"arXiv search error: {str(e)}"
|
| 86 |
+
|
| 87 |
+
@tool
|
| 88 |
+
def web_search(query: str) -> str:
|
| 89 |
+
"""Search the web using DuckDuckGo and return top results."""
|
| 90 |
+
try:
|
| 91 |
+
search = DuckDuckGoSearchRun()
|
| 92 |
+
result = search.run(query)
|
| 93 |
+
return f"Web search results for '{query}':\n{result[:2000]}" # Limit content length
|
| 94 |
+
except Exception as e:
|
| 95 |
+
return f"Web search error: {str(e)}"
|
| 96 |
+
|
| 97 |
+
# --- Load system prompt ---
|
| 98 |
+
with open("system_prompt.txt", "r", encoding="utf-8") as f:
|
| 99 |
+
system_prompt = f.read()
|
| 100 |
+
|
| 101 |
+
# --- Tool Setup ---
|
| 102 |
+
tools = [
|
| 103 |
+
multiply,
|
| 104 |
+
add,
|
| 105 |
+
subtract,
|
| 106 |
+
divide,
|
| 107 |
+
modulus,
|
| 108 |
+
wiki_search,
|
| 109 |
+
arxiv_search,
|
| 110 |
+
web_search,
|
| 111 |
+
]
|
| 112 |
+
|
| 113 |
+
# --- Graph Builder ---
|
| 114 |
+
def build_graph():
|
| 115 |
+
# Initialize model with Gemini 2.5 Flash
|
| 116 |
+
llm = ChatGoogleGenerativeAI(
|
| 117 |
+
model="gemini-2.5-flash",
|
| 118 |
+
temperature=0.3,
|
| 119 |
+
google_api_key=google_api_key,
|
| 120 |
+
max_retries=3
|
| 121 |
+
)
|
| 122 |
+
|
| 123 |
+
# Bind tools to LLM
|
| 124 |
+
llm_with_tools = llm.bind_tools(tools)
|
| 125 |
+
|
| 126 |
+
# 1. 定义状态结构
|
| 127 |
+
class AgentState(TypedDict):
|
| 128 |
+
messages: Annotated[Sequence, operator.add]
|
| 129 |
+
|
| 130 |
+
# 2. 创建图
|
| 131 |
+
workflow = StateGraph(AgentState)
|
| 132 |
+
|
| 133 |
+
# 3. 定义节点函数
|
| 134 |
+
def agent_node(state: AgentState):
|
| 135 |
+
"""主代理节点"""
|
| 136 |
+
try:
|
| 137 |
+
# 添加请求间隔
|
| 138 |
+
time.sleep(1)
|
| 139 |
+
|
| 140 |
+
# 带重试的调用
|
| 141 |
+
@retry(stop=stop_after_attempt(3),
|
| 142 |
+
wait=wait_exponential(multiplier=1, min=4, max=10))
|
| 143 |
+
def invoke_with_retry():
|
| 144 |
+
return llm_with_tools.invoke(state["messages"])
|
| 145 |
+
|
| 146 |
+
response = invoke_with_retry()
|
| 147 |
+
return {"messages": [response]}
|
| 148 |
+
|
| 149 |
+
except Exception as e:
|
| 150 |
+
error_type = "UNKNOWN"
|
| 151 |
+
if "429" in str(e):
|
| 152 |
+
error_type = "QUOTA_EXCEEDED"
|
| 153 |
+
elif "400" in str(e):
|
| 154 |
+
error_type = "INVALID_REQUEST"
|
| 155 |
+
|
| 156 |
+
error_msg = f"AGENT ERROR ({error_type}): {str(e)[:200]}"
|
| 157 |
+
return {"messages": [AIMessage(content=error_msg)]}
|
| 158 |
+
|
| 159 |
+
def tool_node(state: AgentState):
|
| 160 |
+
"""工具执行节点"""
|
| 161 |
+
last_msg = state["messages"][-1]
|
| 162 |
+
tool_calls = last_msg.additional_kwargs.get("tool_calls", [])
|
| 163 |
+
|
| 164 |
+
responses = []
|
| 165 |
+
for call in tool_calls:
|
| 166 |
+
tool_name = call["function"]["name"]
|
| 167 |
+
tool_args = call["function"].get("arguments", {})
|
| 168 |
+
|
| 169 |
+
# 查找工具
|
| 170 |
+
tool_func = next((t for t in tools if t.name == tool_name), None)
|
| 171 |
+
if not tool_func:
|
| 172 |
+
responses.append(f"Tool {tool_name} not available")
|
| 173 |
+
continue
|
| 174 |
+
|
| 175 |
+
try:
|
| 176 |
+
# 解析参数
|
| 177 |
+
if isinstance(tool_args, str):
|
| 178 |
+
tool_args = json.loads(tool_args)
|
| 179 |
+
|
| 180 |
+
# 执行工具
|
| 181 |
+
result = tool_func.invoke(tool_args)
|
| 182 |
+
responses.append(f"{tool_name} result: {result[:1000]}") # 限制结果长度
|
| 183 |
+
except Exception as e:
|
| 184 |
+
responses.append(f"{tool_name} error: {str(e)}")
|
| 185 |
+
|
| 186 |
+
# 修复括号错误:确保正确关闭所有括号
|
| 187 |
+
tool_response_content = "\n".join(responses)
|
| 188 |
+
return {"messages": [AIMessage(content=tool_response_content)]}
|
| 189 |
+
|
| 190 |
+
# 4. 添加节点到工作流
|
| 191 |
+
workflow.add_node("agent", agent_node)
|
| 192 |
+
workflow.add_node("tools", tool_node)
|
| 193 |
+
|
| 194 |
+
# 5. 设置入口点
|
| 195 |
+
workflow.set_entry_point("agent")
|
| 196 |
+
|
| 197 |
+
# 6. 定义条件边
|
| 198 |
+
def should_continue(state: AgentState):
|
| 199 |
+
last_msg = state["messages"][-1]
|
| 200 |
+
|
| 201 |
+
# 错误情况直接结束
|
| 202 |
+
if "AGENT ERROR" in last_msg.content:
|
| 203 |
+
return "end"
|
| 204 |
+
|
| 205 |
+
# 有工具调用则转到工具节点
|
| 206 |
+
if hasattr(last_msg, "tool_calls") and last_msg.tool_calls:
|
| 207 |
+
return "tools"
|
| 208 |
+
|
| 209 |
+
# 包含最终答案则结束
|
| 210 |
+
if "FINAL ANSWER" in last_msg.content:
|
| 211 |
+
return "end"
|
| 212 |
+
|
| 213 |
+
# 其他情况继续代理处理
|
| 214 |
+
return "agent"
|
| 215 |
+
|
| 216 |
+
workflow.add_conditional_edges(
|
| 217 |
+
"agent",
|
| 218 |
+
should_continue,
|
| 219 |
+
{
|
| 220 |
+
"agent": "agent",
|
| 221 |
+
"tools": "tools",
|
| 222 |
+
"end": END
|
| 223 |
+
}
|
| 224 |
+
)
|
| 225 |
+
|
| 226 |
+
# 7. 定义工具节点后的流向
|
| 227 |
+
workflow.add_edge("tools", "agent")
|
| 228 |
+
|
| 229 |
+
# 8. 编译图
|
| 230 |
+
return workflow.compile()
|
| 231 |
+
|
| 232 |
+
# 初始化代理图
|
| 233 |
+
agent_graph = build_graph()
|
system_prompt.txt
ADDED
|
@@ -0,0 +1,32 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
You are GAIA (General AI Assistant). Follow these rules:
|
| 2 |
+
|
| 3 |
+
1. Always use available tools when needed
|
| 4 |
+
2. Format responses as:
|
| 5 |
+
REASONING: <step-by-step logic>
|
| 6 |
+
FINAL ANSWER: <concise result>
|
| 7 |
+
|
| 8 |
+
Available Tools:
|
| 9 |
+
- Math tools: multiply, add, subtract, divide, modulus
|
| 10 |
+
- Search tools:
|
| 11 |
+
* wiki_search (Wikipedia) - for biographical and discography info
|
| 12 |
+
* arxiv_search (arXiv) - for scientific papers
|
| 13 |
+
* web_search (general web search) - for current info
|
| 14 |
+
|
| 15 |
+
For music artist discography queries (like Mercedes Sosa):
|
| 16 |
+
1. Use wiki_search("Artist Name discography") to find the discography section
|
| 17 |
+
2. Identify studio albums specifically (exclude live albums, compilations, etc.)
|
| 18 |
+
3. Check release years carefully - note that reissues or remasters don't count as new studio albums
|
| 19 |
+
4. If Wikipedia is unclear, use web_search for additional sources
|
| 20 |
+
5. Count only albums released within the specified year range
|
| 21 |
+
|
| 22 |
+
Example for Mercedes Sosa question:
|
| 23 |
+
USER: How many studio albums were published by Mercedes Sosa between 2000 and 2009?
|
| 24 |
+
REASONING:
|
| 25 |
+
1. Use wiki_search("Mercedes Sosa discography")
|
| 26 |
+
2. Parse discography section and filter for studio albums
|
| 27 |
+
3. Identify albums released between 2000-2009:
|
| 28 |
+
- Acústico (2002, studio)
|
| 29 |
+
- Corazón Libre (2005, studio)
|
| 30 |
+
- Cantora (2009, studio - recorded in 2009)
|
| 31 |
+
4. Final count: 3 studio albums
|
| 32 |
+
FINAL ANSWER: 3
|