【OLLAMA】LangGrahp调用LLM WIKI
几点说明
LLM WIKI命令
直接将下面的命令给大模型,大模型调用代码执行
将d:/c/1.docx的内容加入项目目录下的knowledge-management/wiki
Ingest d:/c/1.docx into my knowledge-management wiki
Lint 是操作里唯一需要定时跑的,核心风险是幻觉固化:纯 RAG 里一个错误答案就只是错一次;但在 LLM Wiki 里,一个误解会被写进页面,然后悄悄传播到所有关联页面1。Lint 就是定期审计,防止知识系统退化Lint my knowledge-management wiki
项目地址
代码
python
# react_link_agent.py
from datetime import datetime
import re
import json
import subprocess
import os
import sys
from langchain_openai import ChatOpenAI
import requests
from docx import Document as DocxDocument
from langgraph.graph import StateGraph, MessagesState
from langchain_core.messages import AIMessage, SystemMessage, HumanMessage
from langgraph.prebuilt import ToolNode
from langchain_core.tools import tool
SPLIT_DOCS_JSON = r"D:\ollma\faissdb\wf10_faiss_db\split_docs.json"
os.environ.pop("HTTP_PROXY", None)
os.environ.pop("HTTPS_PROXY", None)
os.environ.pop("http_proxy", None)
os.environ.pop("https_proxy", None)
os.environ["NO_PROXY"] = "10.159.188.169,127.0.0.1,localhost,::1"
# ===================== 配置区,自行修改 =====================
OLLAMA_BASE_URL = "http://10.159.188.169:11434"
MODEL_NAME = "qw14b:latest"
SKILL_PATH = r"D:\ollma\ragapp"
MAX_REACT_ROUND = 15 # 防止死循环
# ---------------------- 1. 定义工具 ----------------------
@tool
def load_skill(skill_name: str) -> str:
"""加载 skill_name 对应的 SKILL.md 内容。
Args:
skill_name: 技能名,例如 karpathy-llm-wiki
"""
skill_path = os.path.join(SKILL_PATH, skill_name, "SKILL.md")
if not os.path.exists(skill_path):
return json.dumps({"ok": False, "msg": "技能文件不存在:" + skill_path}, ensure_ascii=False)
with open(skill_path, "r", encoding="utf-8") as f:
return f.read()
@tool
def read_docx(file_path: str) -> str:
"""读取本地 .docx 文件并提取纯文本内容。
Args:
file_path: docx 文件的绝对路径
"""
try:
real_path = file_path
if not os.path.exists(real_path):
return "错误:文件路径不存在 - " + real_path
low = real_path.lower()
if low.endswith((".doc", ".wps", ".rtf", ".odt")):
return "提示:这是旧版二进制格式,read_docx 读不了,请改用 read_office 工具。"
if not low.endswith(".docx"):
return "提示:非 docx 文件,纯文本请用 read_file,旧格式用 read_office。"
doc = DocxDocument(real_path)
parts = []
for para in doc.paragraphs:
if para.text.strip():
parts.append(para.text.strip())
for table in doc.tables:
for row in table.rows:
parts.append(" | ".join([c.text.strip() for c in row.cells]))
full_text = "\n".join(parts)
if not full_text.strip():
return "警告:文件已打开但未提取到任何文本(可能是纯图片扫描件,需 OCR)。"
return "成功读取 " + real_path + "(共 " + str(len(full_text)) + " 字符)\n\n" + full_text
except ValueError as e:
return "路径错误:" + str(e)
except Exception as e:
return "读取 docx 异常: " + str(e)
@tool
def read_file(file_path: str) -> str:
"""读取本地纯文本类文件(txt / md / json / csv / py / log)。
Args:
file_path: 文件的绝对路径
"""
try:
real_path = file_path
if not os.path.exists(real_path):
return "错误:文件路径不存在 - " + real_path
low = real_path.lower()
if low.endswith(".docx"):
return "提示:docx 请改用 read_docx 工具。"
if low.endswith((".doc", ".wps", ".rtf", ".odt")):
return "提示:旧版二进制格式,read_file 读不了,请改用 read_office 工具。"
with open(real_path, "r", encoding="utf-8", errors="ignore") as f:
text = f.read()
if not text.strip():
return "警告:文件内容为空。"
return "成功读取 " + real_path + "(共 " + str(len(text)) + " 字符)\n\n" + text
except ValueError as e:
return "路径错误:" + str(e)
except Exception as e:
return "读取文件异常: " + str(e)
@tool
def save_md(file_path: str, content: str) -> str:
"""
自动建文件夹,将文本内容保存为 Markdown 文件
Args:
file_path: 文件保存路径,例如 "./output.md"
content: markdown 文本内容
return: 结果信息字符串
"""
try:
if not file_path.endswith(".md"):
file_path += ".md"
os.makedirs(os.path.dirname(file_path), exist_ok=True) # 自动建文件夹
with open(file_path, "w", encoding="utf-8") as f:
f.write(content)
return f"✅ Saved: {file_path}"
except Exception as e:
return f"❌ {e}"
tools = [save_md, read_file,read_docx,load_skill]
tool_executor = ToolNode(tools)
# ---------------------- 2. 初始化LLM,绑定工具 ----------------------
llm = ChatOpenAI(
base_url="http://:80/y/llm/v1",
api_key="tyXUvrYHFCa",
model="auto",
temperature=0
)
llm_with_tools = llm.bind_tools(tools)
# ---------------------- 3. 定义节点函数 ----------------------
def call_model(state: MessagesState):
"""agent节点:调用大模型"""
messages = state["messages"]
system_extra = SystemMessage(content=f"""
你是一个基于 Karpathy LLM Wiki 模式的个人知识库维护 Agent。核心职责是"编译并持续维护"知识,而非查询时检索(RAG)。
1. 收到知识库构建需求,第一步调用 load_skill工具,参数skill_name是karpathy-llm-wiki,skill加载成功后,严格按照该skill返回文档里的工作流、任务清单持续执行,使用read_file、save_md完成全部文件操作。
2. 全部任务完成前,持续工具调用,不提前结束、不直接回答用户。
3. 文件修改规则:更新已有md前先read_file读取原内容,追加写入,禁止覆盖。
4. wiki页面统一输出到目录:knowledge-management/wiki/
5. raw源文件统一输出到目录:knowledge-management/raw/llm-wiki/
6. index.md、shturl 在 knowledge-management/wiki/下
""")
new_messages = [system_extra] + messages
# 非流式
response = llm_with_tools.invoke(new_messages)
print("response",response)
# 返回新消息,框架自动追加到state["messages"]
# ========== 打印思考过程 ==========
print("\n🤖【模型本轮思考】")
print(f"思考文本(content):{repr(response.content)}")
print(f"工具调用(tool_calls):{response.tool_calls}")
print("-" * 60)
return {"messages": [response]}
# ---------------------- 4. 条件路由:判断是否继续调用工具 ----------------------
def should_continue(state: MessagesState):
last_msg = state["messages"][-1]
if isinstance(last_msg, AIMessage) and last_msg.tool_calls:
print("✅ 识别到tool_calls,跳转tools节点")
return "tools"
print("✅ 无工具调用,图结束,返回最终回答")
return "__end__"
# ---------------------- 5. 构图 ----------------------
builder = StateGraph(MessagesState)
# 注册节点
builder.add_node("agent", call_model)
builder.add_node("tools", tool_executor)
builder.add_edge("__start__", "agent")
# 条件边:agent执行完,走should_continue判断
builder.add_conditional_edges(
"agent",
should_continue,
)
# 固定边:工具执行完,回到agent
builder.add_edge("tools", "agent")
# 入口节点
builder.set_entry_point("agent")
# 编译图
graph = builder.compile()
if __name__ == "__main__":
print("✅ llm wiki start\n")
while True:
q = input("❓问题:")
if q.lower() == "exit":
break
start = datetime.now()
res1 = graph.invoke({
"messages": [HumanMessage(content=q)]
})
answer=res1["messages"][-1].content
print("\n💡Agent最终回答:")
print(answer)
cost = (datetime.now() - start).total_seconds()
print("-"*70)