What are Vector Embeddings and Vector Search in Azure Cache for Redis

  • Article

Vector similarity search (VSS) has become a popular technology for AI-powered intelligent applications. Azure Cache for Redis can be used as a vector database by combining it models like Azure OpenAI for Retrieval-Augmented Generative AI and analysis scenarios. This article is a high-level introduction to the concept of vector embeddings, vector similarity search, and how Redis can be used as a vector database powering intelligent applications.

For tutorials and sample applications on how to use Azure Cache for Redis and Azure OpenAI to perform vector similarity search, see the following:

Scope of Availability

Vector search capabilities in Redis require Redis Stack, specifically the RediSearch module. This capability is only available in the Enterprise tiers of Azure Cache for Redis.

This table contains the information for vector search availability in different tiers.

Tier Basic / Standard Premium Enterprise Enterprise Flash
Available No No Yes Yes (preview)

What are vector embeddings?

Concept

Vector embeddings are a fundamental concept in machine learning and natural language processing that enable the representation of data, such as words, documents, or images as numerical vectors in a high-dimension vector space. The primary idea behind vector embeddings is to capture the underlying relationships and semantics of the data by mapping them to points in this vector space. That means converting your text or images into a sequence of numbers that represents the data, and then comparing the different number sequences. This allows complex data to be manipulated and analyzed mathematically, making it easier to perform tasks like similarity comparison, recommendation, and classification.

Each machine learning model classifies data and produces the vector in a different manner. Furthermore, it's typically not possible to determine exactly what semantic meaning each vector dimension represents. But because the model is consistent between each block of input data, similar words, documents, or images have vectors that are also similar. For example, the words basketball and baseball have embeddings vectors much closer to each other than a word like rainforest.

Vector comparison

Vectors can be compared using various metrics. The most popular way to compare vectors is to use cosine similarity, which measures the cosine of the angle between two vectors in a multi-dimensional space. The closer the vectors, the smaller the angle. Other common distance metrics include Euclidean distance and inner product.

Generating embeddings

Many machine learning models support embeddings APIs. For an example of how to create vector embeddings using Azure OpenAI Service, see Learn how to generate embeddings with Azure OpenAI.

What is a vector database?

A vector database is a database that can store, manage, retrieve, and compare vectors. Vector databases must be able to efficiently store a high-dimensional vector and retrieve it with minimal latency and high throughput. Nonrelational datastores are most commonly used as vector databases, although it's possible to use relational databases like PostgreSQL, for example, with the pgvector extension.

Index and search method

Vector databases need to index data for fast search and retrieval. In addition, a vector database should support built-in search queries for simplified programming experiences.

There are several indexing methods, such as:

  • FLAT - Brute-force index
  • HNSW - Efficient and robust approximate nearest neighbor search using Hierarchical Navigable Small World graphs

There are several common search methods, including:

  • K-Nearest Neighbors (KNN) - an exhaustive method that provides the most precision but with higher computational cost.
  • Approximate Nearest Neighbors (ANN) - a more efficient by trading precision for greater speed and lower processing overhead.

Search capabilities

Finally, vector databases execute vector searches by using the chosen vector comparison method to return the most similar vectors. Some vector databases can also perform hybrid searches by first narrowing results based on characteristics or metadata also stored in the database before conducting the vector search. This is a way to make the vector search more effective and customizable. For example, a vector search could be limited to only vectors with a specific tag in the database, or vectors with geolocation data in a certain region.

Vector search key scenarios

Vector similarity search can be used in multiple applications. Some common use-cases include:

  • Semantic Q&A. Create a chatbot that can respond to questions about your own data. For instance, a chatbot that can respond to employee questions on their healthcare coverage. Hundreds of pages of dense healthcare coverage documentation can be split into chunks, converted into embeddings vectors, and searched based on vector similarity. The resulting documents can then be summarized for employees using another large language model (LLM). Semantic Q&A Example
  • Document Retrieval. Use the deeper semantic understanding of text provided by LLMs to provide a richer document search experience where traditional keyword-based search falls short. Document Retrieval Example
  • Product Recommendation. Find similar products or services to recommend based on past user activities, like search history or previous purchases. Product Recommendation Example
  • Visual Search. Search for products that look similar to a picture taken by a user or a picture of another product. Visual Search Example
  • Semantic Caching. Reduce the cost and latency of LLMs by caching LLM completions. LLM queries are compared using vector similarity. If a new query is similar enough to a previously cached query, the cached query is returned. Semantic Caching example using LangChain
  • LLM Conversation Memory. Persist conversation history with an LLM as embeddings in a vector database. Your application can use vector search to pull relevant history or "memories" into the response from the LLM. LLM Conversation Memory example

Why choose Azure Cache for Redis for storing and searching vectors?

Azure Cache for Redis can be used effectively as a vector database to store embeddings vectors and to perform vector similarity searches. Support for vector storage and search has been available in many key machine learning frameworks like:

These frameworks feature rich integrations with Redis. For example, the Redis LangChain integration automatically generates an index schema for metadata passed in when using Redis as a vector store. This makes it much easier to filter results based on metadata.

Redis has a wide range of search capabilities through the RediSearch module, which is available in the Enterprise tier of Azure Cache for Redis. These include:

  • Multiple distance metrics, including Euclidean, Cosine, and Internal Product.
  • Support for both KNN (using FLAT) and ANN (using HNSW) indexing methods.
  • Vector storage in hash or JSON data structures
  • Top K queries
  • Vector range queries (that is, find all items within a specific vector distance)
  • Hybrid search with powerful query features such as:
    • Geospatial filtering
    • Numeric and text filters
    • Prefix and fuzzy matching
    • Phonetic matching
    • Boolean queries

Additionally, Redis is often an economical choice because it's already so commonly used for caching or session store applications. In these scenarios, it can pull double-duty by serving a typical caching role while simultaneously handling vector search applications.

What are my other options for storing and searching for vectors?

There are multiple other solutions on Azure for vector storage and search. Other solutions include:

Related content

The best way to get started with embeddings and vector search is to try it yourself!

Tutorial: Conduct vector similarity search on Azure OpenAI embeddings using Azure Cache for Redis