Vector database

Introduction

A vector database is a type of database specifically designed for storing and querying high-dimensional vector data, which is often used in artificial intelligence applications (AI apps). Complex data, including unstructured forms like documents, images, videos, and plain text, is growing rapidly. Traditional databases designed for structured data struggle to effectively store and analyze complex data, often requiring extensive keyword and metadata classification. Vector databases address this issue by transforming complex data into vector embeddings, which describe data objects in numerous dimensions. These databases are gaining popularity due to their ability to extend large language models (LLMs) with long-term memory and provide efficient querying for AI-driven applications.

Vector Embeddings

Vectors are arrays of numbers that can represent complex objects such as words, sentences, images, or audio files in a continuous high-dimensional space, called an embedding. Embeddings work by mapping semantically similar words or features from various data types together. These embeddings can be used in recommendation systems, search engines, and text generation applications like ChatGPT.

Database Structure

In a relational database, data is organized in rows and columns, while in a document database, it is organized in documents and collections. In contrast, a vector database stores arrays of numbers that are clustered based on similarity. These databases can be queried with ultra-low latency, making them ideal for AI-driven applications.

Vector Database Products

Several vector databases have emerged to cater to the growing demand for AI applications. Some of the popular native vector databases include open-source options like Weaviate and Milvus, both written in Go. Pinecone is another popular vector database, although it is not open source. Chroma, based on Clickhouse, is an open-source project with a growing following. Relational databases like Postgres have tools like pgVector, and Redis has first-class vector support to accommodate this type of functionality.

Using Vector Databases with Large Language Models

One of the primary reasons for the increasing popularity of vector databases is their ability to extend large language models (LLMs) with long-term memory. By providing a general-purpose model, such as OpenAI's GPT-4, Meta's LLaMA, or Google's LaMDA, users can store their own data in a vector database. When prompted, the database can query relevant documents to update the context, customizing the final response and providing the AI with long-term memory.

In addition, vector databases can integrate with tools like LangChain, which combine multiple LLMs together for more advanced applications.

Why Use a Vector Database?

Vector databases offer several use cases, including:

Semantic search

Unlike lexical search, which relies on exact word or string matches, semantic search uses the meaning and context of a search query or question. Vector databases use Natural Language Processing models to store and index vector embeddings, allowing for more accurate and relevant search results.

Similarity search for unstructured data

Vector databases facilitate search and retrieval of unstructured data like images, audio, video, and JSON, which can be challenging to classify and store in traditional databases.

Ranking and recommendation engines

By finding similar items based on nearest matches, vector databases are suitable for powering ranking and recommendation engines for online retailers and streaming media services.

Deduplication and record matching

Vector similarity search can be used to find near-duplicate records for applications such as removing duplicate items from a catalog

Anomaly detection

Vector databases can identify anomalies in applications used for threat assessment, fraud detection, and IT operations by finding objects that are distant or dissimilar from expected results.

Required Capabilities of a Vector Database

A vector database must possess certain capabilities to be effective:

Vector Indexes for Search and Retrieval

Vector databases must utilize algorithms designed to index and retrieve vectors efficiently. These algorithms can be optimized depending on the use case requirements. Common metrics used in vector indexes include Euclidean distance, cosine similarity, and dot products. Approximate Nearest Neighbor (ANN) search can balance precision and performance by approximating and retrieving the most similar vectors.

Single-Stage Filtering

Single-stage filtering combines the accuracy of pre-filtering with the speed of post-filtering by merging vector and metadata indexes into a single index. This feature is essential for an effective vector database.

Data Sharding

To achieve scalable and cost-effective performance, vector databases must support horizontal scaling through data sharding. By dividing vectors into shards and replicas, vector databases can scale across multiple machines, allowing for the efficient searching of large datasets.

Replication

Replication is a technique used by vector databases to handle multiple requests efficiently. This approach increases the capacity of the system to process simultaneous or rapidly occurring vector search requests.

Shards and Replicas

Sharding enables a vector database to distribute work across several pods, which can perform vector searches more quickly. Replicas, on the other hand, create copies of the entire set of pods to handle more requests concurrently. In a scenario where multiple search requests are incoming, replicas can help maintain system performance by providing additional processing capacity.

High Availability

Replicas also improve the availability of a vector database. As machines are prone to failure, it is crucial for the system to recover from these events quickly. By distributing replicas across different availability zones, the database can achieve high availability and resilience against simultaneous failures. Users also have a responsibility to ensure sufficient replica capacity, so the remaining replicas can maintain acceptable latency during a failure event.

Hybrid Storage

Hybrid storage configurations help address the challenges of memory cost and search latency associated with vector searches. These configurations balance the need for speed and accuracy with the requirement for cost-effective infrastructure.

In-Memory and On-Disk Storage

In a hybrid storage system, a compressed vector index is stored in memory (RAM), while the original, full-resolution vector index is stored on disk. The in-memory index is used to locate a small set of candidates, which are then searched within the complete index on disk. This approach enables rapid and accurate search results while reducing infrastructure costs by up to 10 times.

Improved Storage Capacity

Hybrid storage allows for more vectors to be stored across the same data footprint, thus lowering the operational cost of a vector database. This is achieved by enhancing overall storage capacity without negatively affecting database performance.

API

APIs play a crucial role in allowing developers to interact with vector databases without the need to build and maintain vector search functionality themselves. By utilizing APIs, developers can focus on optimizing their applications.

REST APIs and Language Clients

Vector databases often offer REST APIs, which add flexibility by allowing the database to be accessed from any environment capable of making HTTPS calls. Developers can also interact with the vector database using clients in various programming languages, such as Python, Java, and Go. These APIs enable actions such as upserting vectors, retrieving query results, or deleting vectors to be performed seamlessly within the context of an application.