335 lines
10 KiB
Python
335 lines
10 KiB
Python
#!/usr/bin/env python3
|
||
"""
|
||
生成知识库问答Agent示例
|
||
展示如何使用向量数据库和RAG技术构建知识库问答系统，包含：
|
||
- 文本向量化（HTTP节点调用embedding API）
|
||
- 向量数据库检索（vector_db节点）
|
||
- 基于检索结果的答案生成（LLM节点）
|
||
- 上下文整合和格式化
|
||
"""
|
||
import sys
|
||
import os
|
||
sys.path.insert(0, os.path.dirname(os.path.dirname(os.path.abspath(__file__))))
|
||
|
||
from sqlalchemy.orm import Session
|
||
from app.core.database import SessionLocal
|
||
from app.models.agent import Agent
|
||
from app.models.user import User
|
||
from datetime import datetime
|
||
import uuid
|
||
|
||
|
||
def generate_knowledge_base_qa_agent(db: Session, user: User):
|
||
"""生成知识库问答Agent"""
|
||
nodes = []
|
||
edges = []
|
||
|
||
# ========== 1. 开始节点 ==========
|
||
start_node = {
|
||
"id": "start-1",
|
||
"type": "start",
|
||
"position": {"x": 50, "y": 400},
|
||
"data": {
|
||
"label": "开始",
|
||
"output_format": "json"
|
||
}
|
||
}
|
||
nodes.append(start_node)
|
||
|
||
# ========== 2. 问题预处理节点 ==========
|
||
preprocess_node = {
|
||
"id": "transform-preprocess",
|
||
"type": "transform",
|
||
"position": {"x": 250, "y": 400},
|
||
"data": {
|
||
"label": "问题预处理",
|
||
"mode": "merge",
|
||
"mapping": {
|
||
"query": "{{query}}",
|
||
"user_id": "{{user_id}}",
|
||
"timestamp": "{{timestamp}}"
|
||
}
|
||
}
|
||
}
|
||
nodes.append(preprocess_node)
|
||
|
||
# ========== 3. 文本向量化节点（HTTP调用embedding API）==========
|
||
# 注意：需要在HTTP节点配置中手动设置Authorization header，使用环境变量DEEPSEEK_API_KEY
|
||
embedding_node = {
|
||
"id": "http-embedding",
|
||
"type": "http",
|
||
"position": {"x": 450, "y": 400},
|
||
"data": {
|
||
"label": "文本向量化",
|
||
"method": "POST",
|
||
"url": "https://api.deepseek.com/v1/embeddings",
|
||
"headers": {
|
||
"Content-Type": "application/json",
|
||
"Authorization": "Bearer sk-fdf7cc1c73504e628ec0119b7e11b8cc"
|
||
},
|
||
"body": {
|
||
"model": "deepseek-embedding",
|
||
"input": "{{query}}"
|
||
},
|
||
"response_format": "json"
|
||
}
|
||
}
|
||
nodes.append(embedding_node)
|
||
|
||
# ========== 4. 提取embedding向量节点 ==========
|
||
extract_embedding_node = {
|
||
"id": "json-extract-embedding",
|
||
"type": "json",
|
||
"position": {"x": 650, "y": 400},
|
||
"data": {
|
||
"label": "提取向量",
|
||
"operation": "extract",
|
||
"path": "output.data.data[0].embedding"
|
||
}
|
||
}
|
||
nodes.append(extract_embedding_node)
|
||
|
||
# ========== 5. 准备向量搜索数据节点 ==========
|
||
prepare_search_node = {
|
||
"id": "transform-prepare-search",
|
||
"type": "transform",
|
||
"position": {"x": 850, "y": 400},
|
||
"data": {
|
||
"label": "准备搜索数据",
|
||
"mode": "merge",
|
||
"mapping": {
|
||
"embedding": "{{output}}",
|
||
"query": "{{query}}"
|
||
}
|
||
}
|
||
}
|
||
nodes.append(prepare_search_node)
|
||
|
||
# ========== 6. 向量数据库检索节点 ==========
|
||
vector_search_node = {
|
||
"id": "vector-search",
|
||
"type": "vector_db",
|
||
"position": {"x": 1050, "y": 400},
|
||
"data": {
|
||
"label": "知识库检索",
|
||
"operation": "search",
|
||
"collection": "knowledge_base",
|
||
"query_vector": "{{embedding}}",
|
||
"top_k": 5
|
||
}
|
||
}
|
||
nodes.append(vector_search_node)
|
||
|
||
# ========== 7. 整理检索结果节点 ==========
|
||
format_results_node = {
|
||
"id": "transform-format-results",
|
||
"type": "transform",
|
||
"position": {"x": 1250, "y": 400},
|
||
"data": {
|
||
"label": "整理检索结果",
|
||
"mode": "merge",
|
||
"mapping": {
|
||
"query": "{{query}}",
|
||
"search_results": "{{output}}"
|
||
}
|
||
}
|
||
}
|
||
nodes.append(format_results_node)
|
||
|
||
# ========== 8. 生成答案节点（LLM）==========
|
||
answer_node = {
|
||
"id": "llm-answer",
|
||
"type": "llm",
|
||
"position": {"x": 1650, "y": 400},
|
||
"data": {
|
||
"label": "生成答案",
|
||
"provider": "deepseek",
|
||
"model": "deepseek-chat",
|
||
"temperature": "0.7",
|
||
"max_tokens": "2000",
|
||
"prompt": """你是一个专业的知识库问答助手。请基于提供的知识库内容回答用户的问题。
|
||
|
||
用户问题：{{query}}
|
||
|
||
相关知识库内容（从向量搜索中检索到的相关文档）：
|
||
{{search_results}}
|
||
|
||
请根据以上知识库内容回答用户的问题。要求：
|
||
1. 答案要准确、完整，基于知识库内容
|
||
2. 如果知识库中没有相关信息，请明确说明"根据知识库，未找到相关信息"
|
||
3. 答案要清晰、有条理，使用Markdown格式
|
||
4. 如果知识库内容与问题不完全匹配，可以结合常识进行补充说明，但要标注哪些是知识库内容，哪些是补充说明
|
||
|
||
请直接输出答案，不要包含其他格式说明。"""
|
||
}
|
||
}
|
||
nodes.append(answer_node)
|
||
|
||
# ========== 9. 提取最终答案节点 ==========
|
||
extract_answer_node = {
|
||
"id": "json-extract-answer",
|
||
"type": "json",
|
||
"position": {"x": 1850, "y": 400},
|
||
"data": {
|
||
"label": "提取最终答案",
|
||
"operation": "extract",
|
||
"path": "output"
|
||
}
|
||
}
|
||
nodes.append(extract_answer_node)
|
||
|
||
# ========== 10. 结束节点 ==========
|
||
end_node = {
|
||
"id": "end-1",
|
||
"type": "end",
|
||
"position": {"x": 2050, "y": 400},
|
||
"data": {
|
||
"label": "结束"
|
||
}
|
||
}
|
||
nodes.append(end_node)
|
||
|
||
# ========== 连接边 ==========
|
||
edges.append({
|
||
"id": "e1",
|
||
"source": "start-1",
|
||
"target": "transform-preprocess",
|
||
"sourceHandle": "right",
|
||
"targetHandle": "left"
|
||
})
|
||
|
||
edges.append({
|
||
"id": "e2",
|
||
"source": "transform-preprocess",
|
||
"target": "http-embedding",
|
||
"sourceHandle": "right",
|
||
"targetHandle": "left"
|
||
})
|
||
|
||
edges.append({
|
||
"id": "e3",
|
||
"source": "http-embedding",
|
||
"target": "json-extract-embedding",
|
||
"sourceHandle": "right",
|
||
"targetHandle": "left"
|
||
})
|
||
|
||
edges.append({
|
||
"id": "e4",
|
||
"source": "json-extract-embedding",
|
||
"target": "transform-prepare-search",
|
||
"sourceHandle": "right",
|
||
"targetHandle": "left"
|
||
})
|
||
|
||
edges.append({
|
||
"id": "e5",
|
||
"source": "transform-prepare-search",
|
||
"target": "vector-search",
|
||
"sourceHandle": "right",
|
||
"targetHandle": "left"
|
||
})
|
||
|
||
edges.append({
|
||
"id": "e6",
|
||
"source": "vector-search",
|
||
"target": "transform-format-results",
|
||
"sourceHandle": "right",
|
||
"targetHandle": "left"
|
||
})
|
||
|
||
edges.append({
|
||
"id": "e7",
|
||
"source": "transform-format-results",
|
||
"target": "llm-answer",
|
||
"sourceHandle": "right",
|
||
"targetHandle": "left"
|
||
})
|
||
|
||
edges.append({
|
||
"id": "e8",
|
||
"source": "llm-answer",
|
||
"target": "json-extract-answer",
|
||
"sourceHandle": "right",
|
||
"targetHandle": "left"
|
||
})
|
||
|
||
edges.append({
|
||
"id": "e9",
|
||
"source": "json-extract-answer",
|
||
"target": "end-1",
|
||
"sourceHandle": "right",
|
||
"targetHandle": "left"
|
||
})
|
||
|
||
return {
|
||
"name": "知识库问答助手",
|
||
"description": "基于向量数据库和RAG技术的知识库问答系统，支持语义搜索和智能回答。需要先使用向量数据库节点将知识库文档向量化并存储。",
|
||
"workflow_config": {"nodes": nodes, "edges": edges}
|
||
}
|
||
|
||
|
||
def main():
|
||
"""主函数"""
|
||
db = SessionLocal()
|
||
try:
|
||
# 获取或创建测试用户
|
||
user = db.query(User).first()
|
||
if not user:
|
||
print("❌ 未找到用户，请先创建用户")
|
||
return
|
||
|
||
print(f"📝 使用用户: {user.username} (ID: {user.id})")
|
||
|
||
# 生成Agent数据
|
||
agent_data = generate_knowledge_base_qa_agent(db, user)
|
||
|
||
# 检查是否已存在
|
||
existing = db.query(Agent).filter(
|
||
Agent.name == agent_data["name"],
|
||
Agent.user_id == user.id
|
||
).first()
|
||
|
||
if existing:
|
||
print(f"Agent '{agent_data['name']}' 已存在，跳过创建")
|
||
return
|
||
|
||
# 创建Agent
|
||
agent = Agent(
|
||
name=agent_data["name"],
|
||
description=agent_data["description"],
|
||
workflow_config=agent_data["workflow_config"],
|
||
user_id=user.id,
|
||
status="draft"
|
||
)
|
||
db.add(agent)
|
||
db.commit()
|
||
db.refresh(agent)
|
||
|
||
print(f"✅ 成功创建Agent: {agent.name} (ID: {agent.id})")
|
||
print(f" 节点数量: {len(agent_data['workflow_config']['nodes'])}")
|
||
print(f" 连接数量: {len(agent_data['workflow_config']['edges'])}")
|
||
print(f"\n📝 使用说明：")
|
||
print(f" 1. 在Agent管理页面找到 '{agent.name}'")
|
||
print(f" 2. 点击'设计'按钮进入工作流编辑器")
|
||
print(f" 3. 配置HTTP节点的API密钥（DeepSeek API Key）")
|
||
print(f" 4. 使用向量数据库节点将知识库文档向量化并存储到 'knowledge_base' 集合")
|
||
print(f" 5. 点击'发布'按钮发布Agent")
|
||
print(f" 6. 点击'使用'按钮测试问答功能")
|
||
print(f"\n💡 提示：")
|
||
print(f" - 知识库文档需要先通过向量数据库节点的 'upsert' 操作存储")
|
||
print(f" - 每个文档需要包含 'text' 和 'embedding' 字段")
|
||
print(f" - 可以使用HTTP节点调用embedding API将文档文本转为向量")
|
||
|
||
except Exception as e:
|
||
print(f"❌ 创建Agent失败: {str(e)}")
|
||
import traceback
|
||
traceback.print_exc()
|
||
db.rollback()
|
||
finally:
|
||
db.close()
|
||
|
||
|
||
if __name__ == "__main__":
|
||
main()
|