150 lines
5.7 KiB
Python
150 lines
5.7 KiB
Python
import asyncio
|
||
from typing import List, Optional
|
||
|
||
from langchain_core.output_parsers import StrOutputParser
|
||
from langchain_core.prompts import ChatPromptTemplate
|
||
from llama_index.core import Document, VectorStoreIndex
|
||
from llama_index.core.vector_stores import MetadataFilters, MetadataFilter, FilterOperator
|
||
from llama_index.vector_stores.milvus import MilvusVectorStore
|
||
from pydantic import BaseModel, Field
|
||
|
||
from app.config.env import env
|
||
from app.utils.llm_utils import create_embeddings, create_llm
|
||
from app.utils.nltk_utils import load_nltk
|
||
|
||
|
||
class KnowledgeQueryParam(BaseModel):
|
||
question: str = Field(..., description="问题文本")
|
||
kb_code: str = Field(..., description="知识库编码")
|
||
top_k: int = Field(default=5, description="返回结果数量")
|
||
|
||
|
||
class MilvusSearchNode(BaseModel):
|
||
id: str = Field(..., description="文档id")
|
||
text: str = Field(..., description="文档内容")
|
||
metadata: dict = Field(..., description="文档元信息")
|
||
score: float = Field(..., description="相似度分数")
|
||
|
||
|
||
class MilvusSearchResponse(BaseModel):
|
||
answer: str = Field(..., description="根据用户问题从检索结果中提取的内容")
|
||
nodes: List[MilvusSearchNode] = Field(..., description="搜索结果")
|
||
|
||
|
||
class MilvusService:
|
||
def __init__(self):
|
||
milvus_vector_store: Optional[MilvusVectorStore] = None
|
||
self._milvus_vector_store = milvus_vector_store
|
||
|
||
self.embeddings = create_embeddings()
|
||
|
||
# 获取一个MilvusVectorStore实例，如果不存在就异步创建
|
||
async def get_vector_store(self):
|
||
if not self._milvus_vector_store:
|
||
print("create new_vector_store...")
|
||
self._milvus_vector_store = MilvusVectorStore(
|
||
uri=env.milvus_uri,
|
||
user=env.milvus_username,
|
||
password=env.milvus_password,
|
||
|
||
db_name=env.llama_index_database,
|
||
collection_name=env.llama_index_collection,
|
||
dim=env.llama_index_dimension,
|
||
|
||
embedding_field="embedding",
|
||
id_field="id",
|
||
similarity_metric="COSINE",
|
||
consistency_level="Strong",
|
||
overwrite=False,
|
||
)
|
||
return self._milvus_vector_store
|
||
|
||
# 检查Milvus连接是否正常
|
||
async def check_milvus_connection(self):
|
||
await self.async_search(KnowledgeQueryParam(question="hello", kb_code=""))
|
||
print("✅ Milvus connection successful：", f"Milvus://{env.milvus_username}:{env.milvus_password}@{env.milvus_uri}/{env.llama_index_database}/{env.llama_index_collection}")
|
||
|
||
async def async_create_index_from_documents(self, documents: List[Document]) -> VectorStoreIndex:
|
||
vector_store = await self.get_vector_store()
|
||
"""异步创建向量索引并插入文档"""
|
||
vector_index: VectorStoreIndex = VectorStoreIndex.from_vector_store(
|
||
vector_store=vector_store,
|
||
embed_model=self.embeddings,
|
||
)
|
||
|
||
# 预先加载NLTK语料库以避免多线程环境中的竞争条件
|
||
# 否则在多线程环境下，parser.get_nodes_from_documents 可能会出现报错信息：'WordListCorpusReader' object has no attribute '_LazyCorpusLoader__args'
|
||
load_nltk()
|
||
|
||
from llama_index.core.node_parser import SimpleNodeParser
|
||
# parser = HierarchicalNodeParser.from_defaults(chunk_sizes=[2048, 512, 128])
|
||
parser = SimpleNodeParser()
|
||
# nodes = await parser.aget_nodes_from_documents(documents)
|
||
# 使用同步方法并包装在 asyncio.to_thread 中
|
||
nodes = await asyncio.to_thread(parser.get_nodes_from_documents, documents)
|
||
await vector_index.ainsert_nodes(nodes)
|
||
return vector_index
|
||
|
||
# 检索milvus文档
|
||
async def async_search(self, param: KnowledgeQueryParam) -> MilvusSearchResponse:
|
||
"""异步向量搜索"""
|
||
# 创建查询引擎，设置top_k参数
|
||
vector_index = VectorStoreIndex.from_vector_store(
|
||
vector_store=await self.get_vector_store(),
|
||
embed_model=self.embeddings,
|
||
)
|
||
|
||
filters = MetadataFilters(filters=[MetadataFilter(key="parent_code", value=param.kb_code, operator=FilterOperator.EQ)])
|
||
|
||
retriever = vector_index.as_retriever(filters=filters, similarity_top_k=param.top_k)
|
||
|
||
result_nodes = await retriever.aretrieve(param.question)
|
||
|
||
chain = ChatPromptTemplate.from_template("""
|
||
你将收到一个用户问题和一组来自LlamaIndex的检索结果。你的任务是：
|
||
1. 分析检索结果内容
|
||
2. 提取与用户问题直接相关的信息片段
|
||
3. 基于这些相关信息生成准确、简洁且有帮助的回答
|
||
|
||
用户问题如下<question/>标签中的内容所示：
|
||
<question>
|
||
{question}
|
||
</question>
|
||
|
||
检索结果如下<context/>标签中的内容所示：
|
||
<context>
|
||
{context}
|
||
</context>
|
||
|
||
处理要求：
|
||
- 仅关注检索结果中与用户问题直接相关的内容
|
||
- 忽略任何无关或关联性较弱的信息
|
||
- 如果检索结果中没有相关信息，请明确回答："在提供的资料中未找到相关信息"
|
||
- 不要添加检索结果之外的额外知识
|
||
- 回答时优先使用检索结果中的原文表述
|
||
""") | create_llm() | StrOutputParser()
|
||
|
||
chain_input = {"context": "\n".join([node.node.text for node in result_nodes]), "question": param.question}
|
||
|
||
relative_content = await chain.ainvoke(chain_input)
|
||
|
||
return MilvusSearchResponse(
|
||
answer=relative_content,
|
||
nodes=[
|
||
MilvusSearchNode(text=node.node.text, metadata=node.node.metadata, score=node.score)
|
||
for node in result_nodes
|
||
]
|
||
)
|
||
|
||
async def async_delete(self, doc_id: str):
|
||
vector_index = VectorStoreIndex.from_vector_store(
|
||
vector_store=await self.get_vector_store(),
|
||
embed_model=self.embeddings,
|
||
)
|
||
await vector_index.adelete_ref_doc(doc_id)
|
||
return None
|
||
|
||
|
||
# 全局实例
|
||
milvus_service = MilvusService()
|