自学内容网 自学内容网

LLM之milvus的使用记录

前言

试试这个数据库

Milvus安装

python接口:pip install pymilvus

docker安装:

通过docker-compose + xxx.yml文件实现

wget https://github.com/milvus-io/milvus/releases/download/v2.4.0/milvus-standalone-docker-compose.yml -O docker-compose.yml

sudo docker-compose up -d

如果没有安装过docker-compose,会报

执行下面这行代码,再重复上面的代码,当然如果你网络不好,可能就需要考虑添加镜像源啥的了 

sudo apt  install docker-compose

安装完之后,运行 

docker ps

会显示你在运行中的容器,这边安装好之后会出现

Milvus数据类型与python对应的数据类型

MilvusPython
DataType.INT64numpy.int64
DataType.INT32numpy.int32
DataType.INT16numpy.int16
DataType.BOOLbool
DataType.FLOATnumpy.float32
DataType.DOUBLEnumpy.double
DataType.ARRAYlist
DataType.VARCHARstr
DataType.JSONdict
FLOAT_VECTOR(浮点数向量)numpy.ndarray or list (元素为numpy.float)


 

Milvus操作

from pymilvus import connections, FieldSchema, CollectionSchema, DataType, Collection,utility
from pymilvus import MilvusClient
from tqdm import tqdm
from datetime import datetime


class milvus_db():
    def __init__(self,url:str='0.0.0.0',collection_name:str='data_store'):
        # milvus_client = MilvusClient(uri="./milvus_demo.db")
        # collection_name = "my_rag_collection"

        connections.connect(host=url, port="19530")
        #self.delete_collection(collection_name)
        if utility.has_collection(collection_name):
            self.collection  = Collection(name=collection_name)
        else:
            schema = self.get_schema()
            self.collection = Collection(name=collection_name, schema=schema)
        print(self.collection.schema)

    def get_schema(self):
        id = FieldSchema(name="id", dtype=DataType.VARCHAR,max_length=128,is_primary=True)  # 主键索引
        text = FieldSchema(name="text", dtype=DataType.VARCHAR,max_length=58192)
        file_name = FieldSchema(name="file_name", dtype=DataType.VARCHAR,max_length=512)
        text_embedding = FieldSchema(name="text_embedding", dtype=DataType.FLOAT_VECTOR,dim=1024)  # 向量,dim=2代表向量只有两列,自己的数据的话一个向量有多少个元素就多少列
        schema = CollectionSchema(fields=[id, text,file_name,text_embedding], description="文本与文本嵌入存储")  # 描述
        return schema

    def change_collection(self,collection_name):
        schema = self.get_schema()
        self.collection  = Collection(name=collection_name,schema=schema)

    def delete_collection(self,collection_name):
        utility.drop_collection(collection_name)

    def create_index(self,metric_type='L2',index_name='L2'):
        #utility.drop_collection(collection_name=collection_name)
        # self.collection = Collection(name=collection_name, schema=schema)
        index_params = {
            "index_type": "AUTOINDEX",
             "metric_type":metric_type,
            "params": {}
        }
        self.collection.create_index(
            field_name="text_embedding",
            index_params=index_params,
            index_name=index_name
        )
        self.collection.load()

    def drop_index(self):
        self.collection.release()
        self.collection.drop_index()


    def insert_data(self,text_id_list,text_list,file_name_list,text_embedding_list):
        data_list = []
        start = datetime.now()
        for id,text,file_name,text_embedding in zip(text_id_list,text_list,file_name_list,text_embedding_list):
            #data_list.append([[id],[text],[file_name],[text_embedding]])
            self.collection.insert([[id],[text],[file_name],[text_embedding]])
        end = datetime.now()
        print(f'插入数据消化时间{end-start}')

    def search(self,query_embedding, top_k=10,metric_type='L2'):
        search_params = {
            "metric_type": metric_type,
            "params": {"level": 2}
        }

        results = self.collection.search(
            [query_embedding],
            anns_field="text_embedding",
            param=search_params,
            limit=top_k,
            output_fields=["text", "file_name"]
        )

        return results

    def list_collections(self):
        collections_list = utility.list_collections()

        return collections_list


    def reranker_init(self,model_name_or_path,device="cpu"):
        self.reranker = bge_rf = BGERerankFunction(
            model_name=model_name_or_path,  # Specify the model name. Defaults to `BAAI/bge-reranker-v2-m3`.
            device="cpu"  # Specify the device to use, e.g., 'cpu' or 'cuda:0'
        )

    def rereank(self,query,serach_result,top_k,rerank_client=None):
        documents_list = [i.entity.get('text') for i in serach_result[0]]
        #如果外部传入非milvus集成的rerank
        if rerank_client:
            response = rerank_client.rerank(
                query=query,
                documents=documents_list,
                top_n=top_k,
            )
            rerank_results = response['results']
            results = []
            for i in rerank_results:
                index = i['index']
                results.append(serach_result[0][index])


        else:
            results = self.reranker(
                query=query,
                documents=documents_list,
                top_k=top_k,
            )

        return results

Milvus可视化

安装好milvus docker之后,哪怕milvus在运行着都可以继续接下来的步骤哦。

因为输入下面的代码就行

# 执行命令,加个 -d  在后台运行
docker run -d -p 8000:3000 -e MILVUS_URL=127.0.0.1:19530 zilliz/attu:v2.2.8

如果像我一样在服务器上跑着,想在本地电脑上看的话,就在MILVUS_URL输入服务器的ip就行啦,然后attu:v 版本尽量接近你milvus的版本就行

之后输入对应链接就行啦。例如我的http://192.0.0.181:8000

参考链接:Milvus向量数据库基础用法及注意细节

欢迎大家点赞或收藏

大家的点赞或收藏可以鼓励作者加快更新~


原文地址:https://blog.csdn.net/weixin_44598554/article/details/143973443

免责声明:本站文章内容转载自网络资源,如本站内容侵犯了原著者的合法权益,可联系本站删除。更多内容请关注自学内容网(zxcms.com)!