Skip to main content

Redis Vector Store

This notebook covers how to get started with the Redis vector store.

Redis is a popular open-source, in-memory data structure store that can be used as a database, cache, message broker, and queue. It now includes vector similarity search capabilities, making it suitable for use as a vector store.

What is Redis?​

Most developers are familiar with Redis. At its core, Redis is a NoSQL Database in the key-value family that can used as a cache, message broker, stream processing and a primary database. Developers choose Redis because it is fast, has a large ecosystem of client libraries, and has been deployed by major enterprises for years.

On top of these traditional use cases, Redis provides additional capabilities like the Search and Query capability that allows users to create secondary index structures within Redis. This allows Redis to be a Vector Database, at the speed of a cache.

Redis as a Vector Database​

Redis uses compressed, inverted indexes for fast indexing with a low memory footprint. It also supports a number of advanced features such as:

  • Indexing of multiple fields in Redis hashes and JSON
  • Vector similarity search (with HNSW (ANN) or FLAT (KNN))
  • Vector Range Search (e.g. find all vectors within a radius of a query vector)
  • Incremental indexing without performance loss
  • Document ranking (using tf-idf, with optional user-provided weights)
  • Field weighting
  • Complex boolean queries with AND, OR, and NOT operators
  • Prefix matching, fuzzy matching, and exact-phrase queries
  • Support for double-metaphone phonetic matching
  • Auto-complete suggestions (with fuzzy prefix suggestions)
  • Stemming-based query expansion in many languages (using Snowball)
  • Support for Chinese-language tokenization and querying (using Friso)
  • Numeric filters and ranges
  • Geospatial searches using Redis geospatial indexing
  • A powerful aggregations engine
  • Supports for all utf-8 encoded text
  • Retrieve full documents, selected fields, or only the document IDs
  • Sorting results (for example, by creation date)

Clients​

Since Redis is much more than just a vector database, there are often use cases that demand the usage of a Redis client besides just the LangChain integration. You can use any standard Redis client library to run Search and Query commands, but it's easiest to use a library that wraps the Search and Query API. Below are a few examples, but you can find more client libraries here.

ProjectLanguageLicenseAuthorStars
jedisJavaMITRedisStars
redisvlPythonMITRedisStars
redis-pyPythonMITRedisStars
node-redisNode.jsMITRedisStars
nredisstack.NETMITRedisStars

Deployment options​

There are many ways to deploy Redis with RediSearch. The easiest way to get started is to use Docker, but there are are many potential options for deployment such as

Redis connection Url schemas​

Valid Redis Url schemas are:

  1. redis:// - Connection to Redis standalone, unencrypted
  2. rediss:// - Connection to Redis standalone, with TLS encryption
  3. redis+sentinel:// - Connection to Redis server via Redis Sentinel, unencrypted
  4. rediss+sentinel:// - Connection to Redis server via Redis Sentinel, booth connections with TLS encryption

More information about additional connection parameters can be found in the redis-py documentation.

Setup​

To use the RedisVectorStore, you'll need to install the langchain-redis partner package, as well as the other packages used throughout this notebook.

%pip install -qU langchain-redis langchain-huggingface sentence-transformers scikit-learn
Note: you may need to restart the kernel to use updated packages.

Credentials​

Redis connection credentials are passed as part of the Redis Connection URL. Redis Connection URLs are versatile and can accommodate various Redis server topologies and authentication methods. These URLs follow a specific format that includes the connection protocol, authentication details, host, port, and database information. The basic structure of a Redis Connection URL is:

[protocol]://[auth]@[host]:[port]/[database]

Where:

  • protocol can be redis for standard connections, rediss for SSL/TLS connections, or redis+sentinel for Sentinel connections.
  • auth includes username and password (if applicable).
  • host is the Redis server hostname or IP address.
  • port is the Redis server port.
  • database is the Redis database number.

Redis Connection URLs support various configurations, including:

  • Standalone Redis servers (with or without authentication)
  • Redis Sentinel setups
  • SSL/TLS encrypted connections
  • Different authentication methods (password-only or username-password)

Below are examples of Redis Connection URLs for different configurations:

# connection to redis standalone at localhost, db 0, no password
redis_url = "redis://localhost:6379"
# connection to host "redis" port 7379 with db 2 and password "secret" (old style authentication scheme without username / pre 6.x)
redis_url = "redis://:secret@redis:7379/2"
# connection to host redis on default port with user "joe", pass "secret" using redis version 6+ ACLs
redis_url = "redis://joe:secret@redis/0"

# connection to sentinel at localhost with default group mymaster and db 0, no password
redis_url = "redis+sentinel://localhost:26379"
# connection to sentinel at host redis with default port 26379 and user "joe" with password "secret" with default group mymaster and db 0
redis_url = "redis+sentinel://joe:secret@redis"
# connection to sentinel, no auth with sentinel monitoring group "zone-1" and database 2
redis_url = "redis+sentinel://redis:26379/zone-1/2"

# connection to redis standalone at localhost, db 0, no password but with TLS support
redis_url = "rediss://localhost:6379"
# connection to redis sentinel at localhost and default port, db 0, no password
# but with TLS support for booth Sentinel and Redis server
redis_url = "rediss+sentinel://localhost"

Launching a Redis Instance with Docker​

To use Redis with LangChain, you need a running Redis instance. You can start one using Docker with:

docker run -d -p 6379:6379 redis/redis-stack:latest

For this example, we'll use a local Redis instance. If you're using a remote instance, you'll need to modify the Redis URL accordingly.

import os

REDIS_URL = os.getenv("REDIS_URL", "redis://localhost:6379")
print(f"Connecting to Redis at: {REDIS_URL}")
Connecting to Redis at: redis://redis:6379

If you want to get automated tracing of your model calls you can also set your LangSmith API key by uncommenting below:

# os.environ["LANGSMITH_API_KEY"] = getpass.getpass("Enter your LangSmith API key: ")
# os.environ["LANGSMITH_TRACING"] = "true"

Let's check that Redis is up an running by pinging it:

import redis

redis_client = redis.from_url(REDIS_URL)
redis_client.ping()
True

Sample Data​

The 20 newsgroups dataset comprises around 18000 newsgroups posts on 20 topics. We'll use a subset for this demonstration and focus on two categories: 'alt.atheism' and 'sci.space':

from langchain.docstore.document import Document
from sklearn.datasets import fetch_20newsgroups

categories = ["alt.atheism", "sci.space"]
newsgroups = fetch_20newsgroups(
    subset="train", categories=categories, shuffle=True, random_state=42
)

# Use only the first 250 documents
texts = newsgroups.data[:250]
metadata = [
    {"category": newsgroups.target_names[target]} for target in newsgroups.target[:250]
]

len(texts)
API Reference:Document
250

Initialization​

The RedisVectorStore instance can be initialized in several ways:

  • RedisVectorStore.__init__ - Initialize directly
  • RedisVectorStore.from_texts - Initialize from a list of texts (optionally with metadata)
  • RedisVectorStore.from_documents - Initialize from a list of langchain_core.documents.Document objects
  • RedisVectorStore.from_existing_index - Initialize from an existing Redis index

Below we will use the RedisVectorStore.__init__ method using a RedisConfig instance.

pip install -qU langchain-openai
import getpass

os.environ["OPENAI_API_KEY"] = getpass.getpass()

from langchain_openai import OpenAIEmbeddings

embeddings = OpenAIEmbeddings(model="text-embedding-3-large")

We'll use the SentenceTransformer model to create embeddings. This model runs locally and doesn't require an API key.

from langchain_redis import RedisConfig, RedisVectorStore

config = RedisConfig(
    index_name="newsgroups",
    redis_url=REDIS_URL,
    metadata_schema=[
        {"name": "category", "type": "tag"},
    ],
)

vector_store = RedisVectorStore(embeddings, config=config)

Manage vector store​

Add items to vector store​

ids = vector_store.add_texts(texts, metadata)

print(ids[0:10])
['newsgroups:f1e788ee61fe410daa8ef941dd166223', 'newsgroups:80b39032181f4299a359a9aaed6e2401', 'newsgroups:99a3efc1883647afba53d115b49e6e92', 'newsgroups:503a6c07cd71418eb71e11b42589efd7', 'newsgroups:7351210e32d1427bbb3c7426cf93a44f', 'newsgroups:4e79fdf67abe471b8ee98ba0e8a1a055', 'newsgroups:03559a1d574e4f9ca0479d7b3891402e', 'newsgroups:9a1c2a7879b8409a805db72feac03580', 'newsgroups:3578a1e129f5435f9743cf803413f37a', 'newsgroups:9f68baf4d6b04f1683d6b871ce8ad92d']

Let's inspect the first document:

texts[0], metadata[0]
('From: [email protected] (Bill Conner)\nSubject: Re: Not the Omni!\nNntp-Posting-Host: okcforum.osrhe.edu\nOrganization: Okcforum Unix Users Group\nX-Newsreader: TIN [version 1.1 PL6]\nLines: 18\n\nCharley Wingate ([email protected]) wrote:\n: \n: >> Please enlighten me.  How is omnipotence contradictory?\n: \n: >By definition, all that can occur in the universe is governed by the rules\n: >of nature. Thus god cannot break them. Anything that god does must be allowed\n: >in the rules somewhere. Therefore, omnipotence CANNOT exist! It contradicts\n: >the rules of nature.\n: \n: Obviously, an omnipotent god can change the rules.\n\nWhen you say, "By definition", what exactly is being defined;\ncertainly not omnipotence. You seem to be saying that the "rules of\nnature" are pre-existant somehow, that they not only define nature but\nactually cause it. If that\'s what you mean I\'d like to hear your\nfurther thoughts on the question.\n\nBill\n',
 {'category': 'alt.atheism'})

Delete items from vector store​

# Delete documents by passing one or more keys/ids
vector_store.index.drop_keys(ids[0])
1

Inspecting the created Index​

Once the Redis VectorStore object has been constructed, an index will have been created in Redis if it did not already exist. The index can be inspected with both the rvland the redis-cli command line tool. If you installed redisvl above, you can use the rvl command line tool to inspect the index.

# assumes you're running Redis locally (use --host, --port, --password, --username, to change this)
!rvl index listall --port 6379
[32m17:54:50[0m [34m[RedisVL][0m [1;30mINFO[0m   Using Redis address from environment variable, REDIS_URL
[32m17:54:50[0m [34m[RedisVL][0m [1;30mINFO[0m   Indices:
[32m17:54:50[0m [34m[RedisVL][0m [1;30mINFO[0m   1. newsgroups

The Redis VectorStore implementation will attempt to generate index schema (fields for filtering) for any metadata passed through the from_texts, from_texts_return_keys, and from_documents methods. This way, whatever metadata is passed will be indexed into the Redis search index allowing for filtering on those fields.

Below we show what fields were created from the metadata we defined above

!rvl index info -i newsgroups --port 6379
[32m17:54:50[0m [34m[RedisVL][0m [1;30mINFO[0m   Using Redis address from environment variable, REDIS_URL


Index Information:
╭──────────────┬────────────────┬────────────────┬─────────────────┬────────────╮
│ Index Name   │ Storage Type   │ Prefixes       │ Index Options   │   Indexing │
├──────────────┼────────────────┼────────────────┼─────────────────┼────────────┤
│ newsgroups   │ HASH           │ ['newsgroups'] │ []              │          0 │
╰──────────────┴────────────────┴────────────────┴─────────────────┴────────────╯
Index Fields:
╭───────────┬─────────────┬────────┬────────────────┬────────────────┬────────────────┬────────────────┬────────────────┬────────────────┬─────────────────┬────────────────╮
│ Name      │ Attribute   │ Type   │ Field Option   │ Option Value   │ Field Option   │ Option Value   │ Field Option   │   Option Value │ Field Option    │ Option Value   │
├───────────┼─────────────┼────────┼────────────────┼────────────────┼────────────────┼────────────────┼────────────────┼────────────────┼─────────────────┼────────────────┤
│ text      │ text        │ TEXT   │ WEIGHT         │ 1              │                │                │                │                │                 │                │
│ embedding │ embedding   │ VECTOR │ algorithm      │ FLAT           │ data_type      │ FLOAT32        │ dim            │            768 │ distance_metric │ COSINE         │
│ category  │ category    │ TAG    │ SEPARATOR      │ |              │                │                │                │                │                 │                │
╰───────────┴─────────────┴────────┴────────────────┴────────────────┴────────────────┴────────────────┴────────────────┴────────────────┴─────────────────┴────────────────╯
!rvl stats -i newsgroups --port 6379
[32m17:54:51[0m [34m[RedisVL][0m [1;30mINFO[0m   Using Redis address from environment variable, REDIS_URL

Statistics:
╭─────────────────────────────┬────────────╮
│ Stat Key                    │ Value      │
├─────────────────────────────┼────────────┤
│ num_docs                    │ 249        │
│ num_terms                   │ 16178      │
│ max_doc_id                  │ 250        │
│ num_records                 │ 50394      │
│ percent_indexed             │ 1          │
│ hash_indexing_failures      │ 0          │
│ number_of_uses              │ 2          │
│ bytes_per_record_avg        │ 38.2743    │
│ doc_table_size_mb           │ 0.0263586  │
│ inverted_sz_mb              │ 1.83944    │
│ key_table_size_mb           │ 0.00932026 │
│ offset_bits_per_record_avg  │ 10.6699    │
│ offset_vectors_sz_mb        │ 0.089057   │
│ offsets_per_term_avg        │ 1.38937    │
│ records_per_doc_avg         │ 202.386    │
│ sortable_values_size_mb     │ 0          │
│ total_indexing_time         │ 72.444     │
│ total_inverted_index_blocks │ 16207      │
│ vector_index_sz_mb          │ 3.01776    │
╰─────────────────────────────┴────────────╯

Query vector store​

Once your vector store has been created and the relevant documents have been added you will most likely wish to query it during the running of your chain or agent.

Query directly​

Performing a simple similarity search can be done as follows:

query = "Tell me about space exploration"
results = vector_store.similarity_search(query, k=2)

print("Simple Similarity Search Results:")
for doc in results:
    print(f"Content: {doc.page_content[:100]}...")
    print(f"Metadata: {doc.metadata}")
    print()
Simple Similarity Search Results:
Content: From: [email protected] (Terry Ford)
Subject: A flawed propulsion system: Space Shuttle
X-Ad...
Metadata: {'category': 'sci.space'}

Content: From: [email protected]
Subject: Space Design Movies?
Article-I.D.: aurora.1993Apr23.124722.1
...
Metadata: {'category': 'sci.space'}

If you want to execute a similarity search and receive the corresponding scores you can run:

# Similarity search with score and filter
scored_results = vector_store.similarity_search_with_score(query, k=2)

print("Similarity Search with Score Results:")
for doc, score in scored_results:
    print(f"Content: {doc.page_content[:100]}...")
    print(f"Metadata: {doc.metadata}")
    print(f"Score: {score}")
    print()
Similarity Search with Score Results:
Content: From: [email protected] (Terry Ford)
Subject: A flawed propulsion system: Space Shuttle
X-Ad...
Metadata: {'category': 'sci.space'}
Score: 0.569670975208

Content: From: [email protected]
Subject: Space Design Movies?
Article-I.D.: aurora.1993Apr23.124722.1
...
Metadata: {'category': 'sci.space'}
Score: 0.590400338173

Query by turning into retriever​

You can also transform the vector store into a retriever for easier usage in your chains.

retriever = vector_store.as_retriever(search_type="similarity", search_kwargs={"k": 2})
retriever.invoke("What planet in the solar system has the largest number of moons?")
[Document(metadata={'category': 'sci.space'}, page_content='Subject: Re: Comet in Temporary Orbit Around Jupiter?\nFrom: Robert Coe <[email protected]>\nDistribution: world\nOrganization: 1776 Enterprises, Sudbury MA\nLines: 23\n\[email protected] writes:\n\n> >> Also, perihelions of Gehrels3 were:\n> >> \n> >> April  1973     83 jupiter radii\n> >> August 1970     ~3 jupiter radii\n> > \n> > Where 1 Jupiter radius = 71,000 km = 44,000 mi = 0.0005 AU.  So the\n> > 1970 figure seems unlikely to actually be anything but a perijove.\n> > Is that the case for the 1973 figure as well?\n> > -- \n> Sorry, _perijoves_...I\'m not used to talking this language.\n\nHmmmm....  The prefix "peri-" is Greek, not Latin, so it\'s usually used\nwith the Greek form of the name of the body being orbited.  (That\'s why\nit\'s "perihelion" rather than "perisol", "perigee" rather than "periterr",\nand "pericynthion" rather than "perilune".)  So for Jupiter I\'d expect it\nto be something like "perizeon".)   :^)\n\n   ___            _                                           -  Bob\n   /__) _   /    / ) _   _\n(_/__) (_)_(_)  (___(_)_(/_______________________________________ [email protected]\nRobert K. Coe ** 14 Churchill St, Sudbury, Massachusetts 01776 ** 508-443-3265\n'),
 Document(metadata={'category': 'sci.space'}, page_content='From: [email protected] (Dillon Pyron)\nSubject: Re: Why not give $1 billion to first year-long moon residents?\nLines: 42\nNntp-Posting-Host: skndiv.dseg.ti.com\nReply-To: [email protected]\nOrganization: TI/DSEG VAX Support\n\n\nIn article <[email protected]>, [email protected] (Peter Schaefer) writes:\n>In article <[email protected]>, [email protected] writes:\n>|> In article <[email protected]>, [email protected] (Gene Wright) writes:\n>|> > With the continuin talk about the "End of the Space Age" and complaints \n>|> > by government over the large cost, why not try something I read about \n>|> > that might just work.\n>|> > \n>|> > Announce that a reward of $1 billion would go to the first corporation \n>|> > who successfully keeps at least 1 person alive on the moon for a year. \n>|> > Then you\'d see some of the inexpensive but not popular technologies begin \n>|> > to be developed. THere\'d be a different kind of space race then!\n>|> > \n>|> > --\n>|> >   [email protected] (Gene Wright)\n>|> > theporch.raider.net  615/297-7951 The MacInteresteds of Nashville\n>|> ====\n>|> If that were true, I\'d go for it.. I have a few friends who we could pool our\n>|> resources and do it.. Maybe make it a prize kind of liek the "Solar Car Race"\n>|> in Australia..\n>|> Anybody game for a contest!\n>|> \n>|> ==\n>|> Michael Adams, [email protected] -- I\'m not high, just jacked\n>\n>\n>Oh gee, a billion dollars!  That\'d be just about enough to cover the cost of the\n>feasability study!  Happy, Happy, JOY! JOY!\n>\n\nFeasability study??  What a wimp!!  While you are studying, others would be\ndoing.  Too damn many engineers doing way too little engineering.\n\n"He who sits on his arse sits on his fortune"  - Sir Richard Francis Burton\n--\nDillon Pyron                      | The opinions expressed are those of the\nTI/DSEG Lewisville VAX Support    | sender unless otherwise stated.\n(214)462-3556 (when I\'m here)     |\n(214)492-4656 (when I\'m home)     |Texans: Vote NO on Robin Hood.  We need\[email protected]          |solutions, not gestures.\nPADI DM-54909                     |\n\n')]

Usage for retrieval-augmented generation​

For guides on how to use this vector store for retrieval-augmented generation (RAG), see the following sections:

Redis-specific functionality​

Redis offers some unique features for vector search:

Similarity search with metadata filtering​

We can filter our search results based on metadata:

from redisvl.query.filter import Tag

query = "Tell me about space exploration"

# Create a RedisVL filter expression
filter_condition = Tag("category") == "sci.space"

filtered_results = vector_store.similarity_search(query, k=2, filter=filter_condition)

print("Filtered Similarity Search Results:")
for doc in filtered_results:
    print(f"Content: {doc.page_content[:100]}...")
    print(f"Metadata: {doc.metadata}")
    print()
Filtered Similarity Search Results:
Content: From: [email protected] (Terry Ford)
Subject: A flawed propulsion system: Space Shuttle
X-Ad...
Metadata: {'category': 'sci.space'}

Content: From: [email protected]
Subject: Space Design Movies?
Article-I.D.: aurora.1993Apr23.124722.1
...
Metadata: {'category': 'sci.space'}

Maximum marginal relevance search helps in getting diverse results:

# Maximum marginal relevance search with filter
mmr_results = vector_store.max_marginal_relevance_search(
    query, k=2, fetch_k=10, filter=filter_condition
)

print("Maximum Marginal Relevance Search Results:")
for doc in mmr_results:
    print(f"Content: {doc.page_content[:100]}...")
    print(f"Metadata: {doc.metadata}")
    print()
Maximum Marginal Relevance Search Results:
Content: From: [email protected] (Terry Ford)
Subject: A flawed propulsion system: Space Shuttle
X-Ad...
Metadata: {'category': 'sci.space'}

Content: From: [email protected] (Michael Moroney)
Subject: Re: Vulcan? (No, not the guy with the ears!)
...
Metadata: {'category': 'sci.space'}

Chain usage​

The code below shows how to use the vector store as a retriever in a simple RAG chain:

pip install -qU langchain-openai
import getpass
import os

os.environ["OPENAI_API_KEY"] = getpass.getpass()

from langchain_openai import ChatOpenAI

llm = ChatOpenAI(model="gpt-4o-mini")
from langchain_core.output_parsers import StrOutputParser
from langchain_core.prompts import ChatPromptTemplate
from langchain_core.runnables import RunnablePassthrough

# Prompt
prompt = ChatPromptTemplate.from_messages(
    [
        (
            "human",
            """You are an assistant for question-answering tasks. Use the following pieces of retrieved context to answer the question. If you don't know the answer, just say that you don't know. Use three sentences maximum and keep the answer concise.
Question: {question} 
Context: {context} 
Answer:""",
        ),
    ]
)


def format_docs(docs):
    return "\n\n".join(doc.page_content for doc in docs)


rag_chain = (
    {"context": retriever | format_docs, "question": RunnablePassthrough()}
    | prompt
    | llm
    | StrOutputParser()
)

rag_chain.invoke("Describe the Space Shuttle program?")
'The Space Shuttle program was a NASA initiative that enabled reusable spacecraft to transport astronauts and cargo to and from low Earth orbit. It conducted a variety of missions, including satellite deployment, scientific research, and assembly of the International Space Station, and typically carried a crew of five astronauts. Although it achieved many successes, the program faced criticism for its safety concerns and the complexity of its propulsion system.'

Connect to an existing Index​

In order to have the same metadata indexed when using the Redis VectorStore. You will need to have the same index_schema passed in either as a path to a yaml file or as a dictionary. The following shows how to obtain the schema from an index and connect to an existing index.

# write the schema to a yaml file
vector_store.index.schema.to_yaml("redis_schema.yaml")
# now we can connect to our existing index as follows

new_rdvs = RedisVectorStore(
    embeddings,
    redis_url=REDIS_URL,
    schema_path="redis_schema.yaml",
)

results = new_rdvs.similarity_search("Space Shuttle Propulsion System", k=3)
print(results[0])
18:19:58 redisvl.index.index INFO   Index already exists, not overwriting.
page_content='From: [email protected] (Terry Ford)
Subject: A flawed propulsion system: Space Shuttle
X-Added: Forwarded by Space Digest
Organization: [via International Space University]
Original-Sender: [email protected]
Distribution: sci
Lines: 13



For an essay, I am writing about the space shuttle and a need for a better
propulsion system.  Through research, I have found that it is rather clumsy 
(i.e. all the checks/tests before launch), the safety hazards ("sitting
on a hydrogen bomb"), etc..  If you have any beefs about the current
space shuttle program Re: propulsion, please send me your ideas.

Thanks a lot.

--
Terry Ford [[email protected]]
Nepean, Ontario, Canada.
' metadata={'category': 'sci.space'}
# compare the two schemas to verify they are the same
new_rdvs.index.schema == vector_store.index.schema
True

Cleanup vector store​

# Clear vector store
vector_store.index.delete(drop=True)

API reference​

For detailed documentation of all RedisVectorStore features and configurations head to the API reference: https://python.langchain.com/api_reference/redis/vectorstores/langchain_redis.vectorstores.RedisVectorStore.html

Was this page helpful?