Wangzf's Doc
note 2024-09-23 Llm

LLM 框架--Vector Database

向量数据库简介

向量数据库是用于高效计算和管理大量 向量数据 的解决方案。 向量数据库是一种专门用于 存储和检索向量数据(embedding) 的数据库系统。 它与传统的基于关系模型的数据库不同，它主要关注的是 向量数据的特性和相似性。

在向量数据库中，数据被表示为向量形式，每个向量代表一个数据项。 这些向量可以是数字、文本、图像或其他类型的数据。 向量数据库使用高效的 索引 和 查询 算法来加速向量数据的存储和检索过程。

向量数据库原理及优势

向量数据库中的 数据以向量作为基本单位，对向量进行 存储、处理 及 检索。 向量数据库通过计算与目标向量的 余弦距离、点积、Squared L2 等获取与目标向量的相似度。 当处理大量甚至海量的向量数据时，向量数据库 索引 和 查询 算法的效率明显高于传统数据库。

主流向量数据库

  • Chroma：是一个轻量级向量数据库，拥有丰富的功能和简单的 API， 具有简单、易用、轻量的优点，但功能相对简单且不支持 GPU 加速，适合初学者使用。
  • Weaviate：是一个开源向量数据库。 除了支持 相似度搜索 和 最大边际相关性(MMR，Maximal Marginal Relevance)搜索 外， 还可以支持 结合多种搜索算法（基于词法搜索、向量搜索）的混合搜索， 从而搜索提高结果的相关性和准确性。
  • Qdrant：Qdrant 使用 Rust 语言开发， 有 极高的检索效率和 RPS(Requests Per Second)， 支持 本地运行、部署在本地服务器 及 Qdrant 云 三种部署模式。 且可以通过为页面内容和元数据制定不同的键来复用数据。

Chroma

Chroma 简介

Chroma is the AI-native open-source vector database. Chroma makes it easy to build LLM apps by making knowledge, facts, and skills pluggable(可插拔) for LLMs.

img

img

Chroma gives you the tools to:

  • store embeddings and their metadata
  • embed documents and queries
  • search embeddings

Chroma prioritizes:

  • simplicity and developer productivity
  • it also happens to be very quick

Chroma 安装

$ pip insall chromadb

Chroma 使用

  1. 创建一个 Chroma Client
import chromadb

chroma_client = chormadb.Client()
  1. 创建一个 collection

Collections are where you’ll store your embeddings, documents, and any additional metadata.

collection = chroma_client.create_collection(name = "my_collection")
  1. 存储文本文档
collection.add(
    documents = [
        "This is a document about pineapple",
        "This is a document about oranges"
    ],
    ids = ["id1", "id2"],
)
  1. 查询 collection

You can query the collection with a list of query texts, and Chroma will return the n most similar results.

results = collection.query(
    query_texts = ["This is a query document about hawaii"],  # Chroma will embed this for you
    n_results = 2,  # how many results to return
)
print(result)

{
  'documents': [[
      'This is a document about pineapple',
      'This is a document about oranges'
  ]],
  'ids': [['id1', 'id2']],
  'distances': [[1.0404009819030762, 1.243080496788025]],
  'uris': None,
  'data': None,
  'metadatas': [[None, None]],
  'embeddings': None,
}

Chroma API

创建 Chroma Client

import chromadb

client = chromadb.PersistentClient(path = "/path/to/save/to")

# ------------------------------
# useful convenience method
# ------------------------------
# returns a nanosecond heartbeat. 
# Useful for making sure the client remains connected.
client.heartbeat()

# Empties and completely resets the database. 
# ⚠️ This is destructive and not reversible.
client.reset()

启动 Client-Server

本地启动 Chroma Server:

$ chroma run --path /db_path

本地启动 HTTP client:

import chromadb

chroma_client = chromadb.HttpClient(host = "localhost", port = 8000)

本地启动异步(async) HTTP client:

import asyncio
import chromadb

async def main():
    client = await chromadb.AsyncHttpClient()
    collection = await client.create_collection(name = "my_collection")

    await collection.add(
        documents = ["hello world"],
        ids = ["id1"]
    )

asyncio.run(main())

Collections 操作

Creating-Inspecting-Deleting
collection = client.create_collection(name="my_collection", embedding_function=emb_fn)
collection = client.get_collection(name="my_collection", embedding_function=emb_fn)

# Get a collection object from an existing collection, by name. Will raise an exception if it's not found.
collection = client.get_collection(name="test") 
# Get a collection object from an existing collection, by name. If it doesn't exist, create it.
collection = client.get_or_create_collection(name="test") 
# Delete a collection and all associated embeddings, documents, and metadata. 
⚠# ️ This is destructive and not reversible
client.delete_collection(name="my_collection")
修改距离函数

Changing distance function

collection = client.create_collection(
    name = "collection_name",
    metadata = {"hnsw:space": "cosine"} # l2 is the default
)
Add data
collection.add(
    documents = ["lorem ipsum...", "doc2", "doc3", ...],
    metadatas = [
        {"chapter": "3", "verse": "16"}, 
        {"chapter": "3", "verse": "5"}, 
        {"chapter": "29", "verse": "11"}, 
        ...
    ],
    ids = ["id1", "id2", "id3", ...]
)

collection.add(
    documents = ["doc1", "doc2", "doc3", ...],
    embeddings = [
        [1.1, 2.3, 3.2], 
        [4.5, 6.9, 4.4], 
        [1.1, 2.3, 3.2], 
        ...
    ],
    metadatas = [
        {"chapter": "3", "verse": "16"}, 
        {"chapter": "3", "verse": "5"}, 
        {"chapter": "29", "verse": "11"}, 
        ...
    ],
    ids = ["id1", "id2", "id3", ...]
)

collection.add(
    embeddings = [[1.1, 2.3, 3.2], [4.5, 6.9, 4.4], [1.1, 2.3, 3.2], ...],
    metadatas = [
        {"chapter": "3", "verse": "16"}, 
        {"chapter": "3", "verse": "5"}, 
        {"chapter": "29", "verse": "11"}, 
        ...
    ],
    ids = ["id1", "id2", "id3", ...]
)
Query
collection.query(
    query_embeddings = [
        [11.1, 12.1, 13.1],
        [1.1, 2.3, 3.2], 
        ...
    ],
    n_results = 10,
    where = {
        "metadata_field": "is_equal_to_this"
    },
    where_document = {
        "$contains": "search_string"
    },
)
Where

where filters

  • Filtering by metadata
  • Filtering by document content
  • Using logical operators
  • Using inclusion operators
Updating data
collection.update(
    ids = ["id1", "id2", "id3", ...],
    embeddings = [
        [1.1, 2.3, 3.2], 
        [4.5, 6.9, 4.4], 
        [1.1, 2.3, 3.2], 
        ...
    ],
    metadatas = [
        {"chapter": "3", "verse": "16"}, 
        {"chapter": "3", "verse": "5"}, 
        {"chapter": "29", "verse": "11"}, 
        ...
    ],
    documents = ["doc1", "doc2", "doc3", ...],
)

collection.upsert(
    ids = ["id1", "id2", "id3", ...],
    embeddings = [
        [1.1, 2.3, 3.2], 
        [4.5, 6.9, 4.4], 
        [1.1, 2.3, 3.2], 
        ...
    ],
    metadatas = [
        {"chapter": "3", "verse": "16"}, 
        {"chapter": "3", "verse": "5"}, 
        {"chapter": "29", "verse": "11"}, 
        ...
    ],
    documents = ["doc1", "doc2", "doc3", ...],
)
Deleting data
collection.delete(
    ids = [
        "id1", "id2", "id3", ...
    ],
	where = {"chapter": "20"}
)

Chorma 部署

Chroma Server：

  • Client/Server Mode
  • Python Thin-Client

容器(Containers)：

  • Docker
  • Kubernetes

云服务(Cloud Providers)：

  • AWS
  • GCP
  • Azure

Client-Server Mode

Run Chroma in client/server mode by using CLI:

$ chroma run --path /db_path

Connect to Server using Chroma HttpClient：

import chromadb

chroma_client = chromadb.HttpClient(host = 'localhost', port = 8000)

异步运行 AsyncHttpClient：

import asyncio
import chromadb

async def main():
    client = await chromadb.AsyncHttpClient()
    collection = await client.create_collection(name="my_collection")
    await collection.add(
        documents=["hello world"],
        ids=["id1"]
    )

asyncio.run(main())

Chorma’s Thin-Client

$ pip install chromadb-client

import chromadb

# Example setup of the client to connect to your chroma server
client = chromadb.HttpClient(host = 'localhost', port = 8000)

# Or for async usage:
async def main():
    client = await chromadb.AsyncHttpClient(host='localhost', port=8000)

Weaviate

Qdrant

参考