Redis 向量存储

在本笔记本中，我们将快速演示如何使用RedisVectorStore。

如果您在 Colab 上打开这个笔记本，您可能需要安装 LlamaIndex 🦙。

%pip install -U llama-index llama-index-vector-stores-redis llama-index-embeddings-cohere llama-index-embeddings-openai

import os
import getpass
import sys
import logging
import textwrap
import warnings

warnings.filterwarnings("ignore")

# Uncomment to see debug logs
logging.basicConfig(stream=sys.stdout, level=logging.INFO)

from llama_index.core import VectorStoreIndex, SimpleDirectoryReader
from llama_index.vector_stores.redis import RedisVectorStore

启动Redis

启动Redis最简单的方法是使用Redis Stack docker镜像或快速注册一个免费Redis Cloud实例。

要跟随本教程的每一步，请按如下方式启动镜像：

docker run --name redis-vecdb -d -p 6379:6379 -p 8001:8001 redis/redis-stack:latest

这将在端口8001上启动RedisInsight界面，您可以通过http://localhost:8001进行查看。

设置OpenAI

首先让我们添加OpenAI API密钥。这将允许我们访问OpenAI以获取嵌入向量并使用ChatGPT。

oai_api_key = getpass.getpass("OpenAI API Key:")
os.environ["OPENAI_API_KEY"] = oai_api_key

下载数据

!mkdir -p 'data/paul_graham/'
!wget 'https://raw.githubusercontent.com/run-llama/llama_index/main/docs/examples/data/paul_graham/paul_graham_essay.txt' -O 'data/paul_graham/paul_graham_essay.txt'

--2024-04-10 19:35:33--  https://raw.githubusercontent.com/run-llama/llama_index/main/docs/examples/data/paul_graham/paul_graham_essay.txt
Resolving raw.githubusercontent.com (raw.githubusercontent.com)... 2606:50c0:8003::154, 2606:50c0:8000::154, 2606:50c0:8002::154, ...
Connecting to raw.githubusercontent.com (raw.githubusercontent.com)|2606:50c0:8003::154|:443... connected.
HTTP request sent, awaiting response... 200 OK
Length: 75042 (73K) [text/plain]
Saving to: ‘data/paul_graham/paul_graham_essay.txt’

data/paul_graham/pa 100%[===================>]  73.28K  --.-KB/s    in 0.03s

2024-04-10 19:35:33 (2.15 MB/s) - ‘data/paul_graham/paul_graham_essay.txt’ saved [75042/75042]

读取数据集

这里我们将使用一套保罗·格雷厄姆的文章作为文本来源，将其转化为嵌入向量，存储在RedisVectorStore中，并通过查询为我们的LLM问答循环寻找上下文。

# load documents
documents = SimpleDirectoryReader("./data/paul_graham").load_data()
print(
    "Document ID:",
    documents[0].id_,
    "Document Filename:",
    documents[0].metadata["file_name"],
)

Document ID: 7056f7ba-3513-4ef4-9792-2bd28040aaed Document Filename: paul_graham_essay.txt

初始化默认的 Redis 向量存储

现在我们已经准备好了文档，可以使用默认设置初始化Redis向量存储。这将允许我们将向量存储在Redis中，并为实时搜索创建索引。

from llama_index.core import StorageContext
from redis import Redis

# create a Redis client connection
redis_client = Redis.from_url("redis://localhost:6379")

# create the vector store wrapper
vector_store = RedisVectorStore(redis_client=redis_client, overwrite=True)

# load storage context
storage_context = StorageContext.from_defaults(vector_store=vector_store)

# build and load index from documents and storage context
index = VectorStoreIndex.from_documents(
    documents, storage_context=storage_context
)
# index = VectorStoreIndex.from_vector_store(vector_store=vector_store)

19:39:17 llama_index.vector_stores.redis.base INFO   Using default RedisVectorStore schema.
19:39:19 httpx INFO   HTTP Request: POST https://api.openai.com/v1/embeddings "HTTP/1.1 200 OK"
19:39:19 llama_index.vector_stores.redis.base INFO   Added 22 documents to index llama_index

查询默认向量存储

现在我们已经将数据存储在索引中，我们可以对索引进行提问。

该索引将使用数据作为LLM的知识库。as_query_engine()的默认设置使用OpenAI嵌入和GPT作为语言模型。因此，除非您选择自定义或本地语言模型，否则需要OpenAI密钥。

下面我们将测试针对我们索引的搜索，然后使用大型语言模型进行完整的检索增强生成。

query_engine = index.as_query_engine()
retriever = index.as_retriever()

result_nodes = retriever.retrieve("What did the author learn?")
for node in result_nodes:
    print(node)

19:39:22 httpx INFO   HTTP Request: POST https://api.openai.com/v1/embeddings "HTTP/1.1 200 OK"
19:39:22 llama_index.vector_stores.redis.base INFO   Querying index llama_index with filters *
19:39:22 llama_index.vector_stores.redis.base INFO   Found 2 results for query with id ['llama_index/vector_adb6b7ce-49bb-4961-8506-37082c02a389', 'llama_index/vector_e39be1fe-32d0-456e-b211-4efabd191108']
Node ID: adb6b7ce-49bb-4961-8506-37082c02a389
Text: What I Worked On  February 2021  Before college the two main
things I worked on, outside of school, were writing and programming. I
didn't write essays. I wrote what beginning writers were supposed to
write then, and probably still are: short stories. My stories were
awful. They had hardly any plot, just characters with strong feelings,
which I ...
Score:  0.820

Node ID: e39be1fe-32d0-456e-b211-4efabd191108
Text: Except for a few officially anointed thinkers who went to the
right parties in New York, the only people allowed to publish essays
were specialists writing about their specialties. There were so many
essays that had never been written, because there had been no way to
publish them. Now they could be, and I was going to write them. [12]
I've wor...
Score:  0.819

response = query_engine.query("What did the author learn?")
print(textwrap.fill(str(response), 100))

19:39:25 httpx INFO   HTTP Request: POST https://api.openai.com/v1/embeddings "HTTP/1.1 200 OK"
19:39:25 llama_index.vector_stores.redis.base INFO   Querying index llama_index with filters *
19:39:25 llama_index.vector_stores.redis.base INFO   Found 2 results for query with id ['llama_index/vector_adb6b7ce-49bb-4961-8506-37082c02a389', 'llama_index/vector_e39be1fe-32d0-456e-b211-4efabd191108']
19:39:27 httpx INFO   HTTP Request: POST https://api.openai.com/v1/chat/completions "HTTP/1.1 200 OK"
The author learned that working on things that weren't prestigious often led to valuable discoveries
and indicated the right kind of motives. Despite the lack of initial prestige, pursuing such work
could be a sign of genuine potential and appropriate motivations, steering clear of the common
pitfall of being driven solely by the desire to impress others.

result_nodes = retriever.retrieve("What was a hard moment for the author?")
for node in result_nodes:
    print(node)

19:39:27 httpx INFO   HTTP Request: POST https://api.openai.com/v1/embeddings "HTTP/1.1 200 OK"
19:39:27 llama_index.vector_stores.redis.base INFO   Querying index llama_index with filters *
19:39:27 llama_index.vector_stores.redis.base INFO   Found 2 results for query with id ['llama_index/vector_adb6b7ce-49bb-4961-8506-37082c02a389', 'llama_index/vector_e39be1fe-32d0-456e-b211-4efabd191108']
Node ID: adb6b7ce-49bb-4961-8506-37082c02a389
Text: What I Worked On  February 2021  Before college the two main
things I worked on, outside of school, were writing and programming. I
didn't write essays. I wrote what beginning writers were supposed to
write then, and probably still are: short stories. My stories were
awful. They had hardly any plot, just characters with strong feelings,
which I ...
Score:  0.802

Node ID: e39be1fe-32d0-456e-b211-4efabd191108
Text: Except for a few officially anointed thinkers who went to the
right parties in New York, the only people allowed to publish essays
were specialists writing about their specialties. There were so many
essays that had never been written, because there had been no way to
publish them. Now they could be, and I was going to write them. [12]
I've wor...
Score:  0.799

response = query_engine.query("What was a hard moment for the author?")
print(textwrap.fill(str(response), 100))

19:39:29 httpx INFO   HTTP Request: POST https://api.openai.com/v1/embeddings "HTTP/1.1 200 OK"
19:39:29 llama_index.vector_stores.redis.base INFO   Querying index llama_index with filters *
19:39:29 llama_index.vector_stores.redis.base INFO   Found 2 results for query with id ['llama_index/vector_adb6b7ce-49bb-4961-8506-37082c02a389', 'llama_index/vector_e39be1fe-32d0-456e-b211-4efabd191108']
19:39:31 httpx INFO   HTTP Request: POST https://api.openai.com/v1/chat/completions "HTTP/1.1 200 OK"
A hard moment for the author was when one of his programs on the IBM 1401 mainframe didn't
terminate, leading to a technical error and an uncomfortable situation with the data center manager.

index.vector_store.delete_index()

19:39:34 llama_index.vector_stores.redis.base INFO   Deleting index llama_index

使用自定义索引架构

在大多数使用场景中，您需要能够自定义底层索引配置和规范。例如，为了定义您希望启用的特定元数据过滤器，这非常方便。

使用Redis，这就像定义一个索引模式对象（来自文件或字典）并将其传递给向量存储客户端包装器一样简单。

对于这个示例，我们将：

将嵌入模型切换为 Cohere
为文档添加额外的元数据字段 updated_at 时间戳
索引现有的 file_name 元数据字段

from llama_index.core.settings import Settings
from llama_index.embeddings.cohere import CohereEmbedding

# set up Cohere Key
co_api_key = getpass.getpass("Cohere API Key:")
os.environ["CO_API_KEY"] = co_api_key

# set llamaindex to use Cohere embeddings
Settings.embed_model = CohereEmbedding()

from redisvl.schema import IndexSchema


custom_schema = IndexSchema.from_dict(
    {
        # customize basic index specs
        "index": {
            "name": "paul_graham",
            "prefix": "essay",
            "key_separator": ":",
        },
        # customize fields that are indexed
        "fields": [
            # required fields for llamaindex
            {"type": "tag", "name": "id"},
            {"type": "tag", "name": "doc_id"},
            {"type": "text", "name": "text"},
            # custom metadata fields
            {"type": "numeric", "name": "updated_at"},
            {"type": "tag", "name": "file_name"},
            # custom vector field definition for cohere embeddings
            {
                "type": "vector",
                "name": "vector",
                "attrs": {
                    "dims": 1024,
                    "algorithm": "hnsw",
                    "distance_metric": "cosine",
                },
            },
        ],
    }
)

custom_schema.index

IndexInfo(name='paul_graham', prefix='essay', key_separator=':', storage_type=<StorageType.HASH: 'hash'>)

custom_schema.fields

{'id': TagField(name='id', type='tag', path=None, attrs=TagFieldAttributes(sortable=False, separator=',', case_sensitive=False, withsuffixtrie=False)),
 'doc_id': TagField(name='doc_id', type='tag', path=None, attrs=TagFieldAttributes(sortable=False, separator=',', case_sensitive=False, withsuffixtrie=False)),
 'text': TextField(name='text', type='text', path=None, attrs=TextFieldAttributes(sortable=False, weight=1, no_stem=False, withsuffixtrie=False, phonetic_matcher=None)),
 'updated_at': NumericField(name='updated_at', type='numeric', path=None, attrs=NumericFieldAttributes(sortable=False)),
 'file_name': TagField(name='file_name', type='tag', path=None, attrs=TagFieldAttributes(sortable=False, separator=',', case_sensitive=False, withsuffixtrie=False)),
 'vector': HNSWVectorField(name='vector', type='vector', path=None, attrs=HNSWVectorFieldAttributes(dims=1024, algorithm=<VectorIndexAlgorithm.HNSW: 'HNSW'>, datatype=<VectorDataType.FLOAT32: 'FLOAT32'>, distance_metric=<VectorDistanceMetric.COSINE: 'COSINE'>, initial_cap=None, m=16, ef_construction=200, ef_runtime=10, epsilon=0.01))}

了解更多关于模式与索引设计在redis中的应用。

from datetime import datetime


def date_to_timestamp(date_string: str) -> int:
    date_format: str = "%Y-%m-%d"
    return int(datetime.strptime(date_string, date_format).timestamp())


# iterate through documents and add new field
for document in documents:
    document.metadata["updated_at"] = date_to_timestamp(
        document.metadata["last_modified_date"]
    )

vector_store = RedisVectorStore(
    schema=custom_schema,  # provide customized schema
    redis_client=redis_client,
    overwrite=True,
)

storage_context = StorageContext.from_defaults(vector_store=vector_store)

# build and load index from documents and storage context
index = VectorStoreIndex.from_documents(
    documents, storage_context=storage_context
)

19:40:05 httpx INFO   HTTP Request: POST https://api.cohere.ai/v1/embed "HTTP/1.1 200 OK"
19:40:06 httpx INFO   HTTP Request: POST https://api.cohere.ai/v1/embed "HTTP/1.1 200 OK"
19:40:06 httpx INFO   HTTP Request: POST https://api.cohere.ai/v1/embed "HTTP/1.1 200 OK"
19:40:06 llama_index.vector_stores.redis.base INFO   Added 22 documents to index paul_graham

查询向量存储并根据元数据进行筛选

既然我们已经在Redis中索引了额外的元数据，现在让我们尝试一些带过滤器的查询。

from llama_index.core.vector_stores import (
    MetadataFilters,
    MetadataFilter,
    ExactMatchFilter,
)

retriever = index.as_retriever(
    similarity_top_k=3,
    filters=MetadataFilters(
        filters=[
            ExactMatchFilter(key="file_name", value="paul_graham_essay.txt"),
            MetadataFilter(
                key="updated_at",
                value=date_to_timestamp("2023-01-01"),
                operator=">=",
            ),
            MetadataFilter(
                key="text",
                value="learn",
                operator="text_match",
            ),
        ],
        condition="and",
    ),
)

result_nodes = retriever.retrieve("What did the author learn?")

for node in result_nodes:
    print(node)

19:40:22 httpx INFO   HTTP Request: POST https://api.cohere.ai/v1/embed "HTTP/1.1 200 OK"


19:40:22 llama_index.vector_stores.redis.base INFO   Querying index paul_graham with filters ((@file_name:{paul_graham_essay\.txt} @updated_at:[1672549200 +inf]) @text:(learn))
19:40:22 llama_index.vector_stores.redis.base INFO   Found 3 results for query with id ['essay:0df3b734-ecdb-438e-8c90-f21a8c80f552', 'essay:01108c0d-140b-4dcc-b581-c38b7df9251e', 'essay:ced36463-ac36-46b0-b2d7-935c1b38b781']
Node ID: 0df3b734-ecdb-438e-8c90-f21a8c80f552
Text: All that seemed left for philosophy were edge cases that people
in other fields felt could safely be ignored.  I couldn't have put
this into words when I was 18. All I knew at the time was that I kept
taking philosophy courses and they kept being boring. So I decided to
switch to AI.  AI was in the air in the mid 1980s, but there were two
things...
Score:  0.410

Node ID: 01108c0d-140b-4dcc-b581-c38b7df9251e
Text: It was not, in fact, simply a matter of teaching SHRDLU more
words. That whole way of doing AI, with explicit data structures
representing concepts, was not going to work. Its brokenness did, as
so often happens, generate a lot of opportunities to write papers
about various band-aids that could be applied to it, but it was never
going to get us ...
Score:  0.390

Node ID: ced36463-ac36-46b0-b2d7-935c1b38b781
Text: Grad students could take classes in any department, and my
advisor, Tom Cheatham, was very easy going. If he even knew about the
strange classes I was taking, he never said anything.  So now I was in
a PhD program in computer science, yet planning to be an artist, yet
also genuinely in love with Lisp hacking and working away at On Lisp.
In other...
Score:  0.389

从Redis中的现有索引恢复

从索引恢复需要一个Redis连接客户端（或URL）、overwrite=False，并传入之前使用的相同模式对象。（为方便起见，可以使用.to_yaml()将其卸载到YAML文件中）

custom_schema.to_yaml("paul_graham.yaml")

vector_store = RedisVectorStore(
    schema=IndexSchema.from_yaml("paul_graham.yaml"),
    redis_client=redis_client,
)
index = VectorStoreIndex.from_vector_store(vector_store=vector_store)

19:40:28 redisvl.index.index INFO   Index already exists, not overwriting.

在不久的将来 — 我们将实现一个便捷方法，仅通过索引名称即可加载：

RedisVectorStore.from_existing_index(index_name="paul_graham", redis_client=redis_client)

完全删除文档或索引

有时删除文档或整个索引可能很有用。这可以通过使用 delete 和 delete_index 方法来实现。

document_id = documents[0].doc_id
document_id

'7056f7ba-3513-4ef4-9792-2bd28040aaed'

print("Number of documents before deleting", redis_client.dbsize())
vector_store.delete(document_id)
print("Number of documents after deleting", redis_client.dbsize())

Number of documents before deleting 22
19:40:32 llama_index.vector_stores.redis.base INFO   Deleted 22 documents from index paul_graham
Number of documents after deleting 0

然而，Redis索引仍然存在（没有关联文档）以便持续更新插入。

vector_store.index_exists()

True

# now lets delete the index entirely
# this will delete all the documents and the index
vector_store.delete_index()

19:40:37 llama_index.vector_stores.redis.base INFO   Deleting index paul_graham

print("Number of documents after deleting", redis_client.dbsize())

Number of documents after deleting 0

故障排除

如果查询结果为空，有几个问题需要检查：

架构

与其他向量存储不同，Redis 要求用户显式定义索引的架构。这主要有以下几个原因：

Redis 被用于多种应用场景，包括实时向量搜索，同时也用于标准文档存储/检索、缓存、消息传递、发布/订阅、会话管理等。并非记录中的所有属性都需要建立索引以供搜索。这在一定程度上是出于效率考虑，部分也是为了尽量减少用户误操作的可能性。
所有索引模式，在使用Redis和LlamaIndex时，必须至少包含以下字段：id、doc_id、text和vector。

使用默认模式（假设为 OpenAI 嵌入）或自定义模式（见上文）实例化您的 RedisVectorStore。

前缀问题

Redis 期望所有记录都拥有一个键前缀，将键空间划分为“分区”，以便用于可能不同的应用程序、用例和客户端。

请确保所选的 prefix 作为索引模式的一部分，在您的代码中保持一致（与特定索引绑定）。

要查看您的索引是使用什么前缀创建的，可以在 Redis CLI 中运行 FT.INFO <name of your index> 并查看 index_definition => prefixes 下的内容。

数据与索引

Redis 将数据集中的记录和索引视为不同的实体。这使您在执行更新、更新插入和索引模式迁移时具有更大的灵活性。

如果您已有现有索引并希望确保它被删除，可以在Redis CLI中运行FT.DROPINDEX <name of your index>。请注意，除非您传递DD，否则这不会删除您的实际数据。

使用元数据时的空查询

如果你在索引已经创建后添加元数据，然后尝试基于该元数据进行查询，你的查询将返回空结果。

Redis仅在创建索引时对字段进行索引（类似于上述处理前缀的方式）。