A Brief Comparison of Vector Databases

Jayita Bhattacharyya
CodeX
Published in
3 min readMay 6, 2024

In this supercharged era of RAG (Retrieval Augmented Generation) fueled by LLMs (Large Language Models), let's look at the roots or surface of these applications built on top of the mighty VD (Vector Databases). In the cluttered market of Vector Databases which is flooding day by day, it's hard to decide which one to use and enhance application performance based on its features and capabilities. Picking up the Vector database relies on factors like open-source vs closed-source, on-premise or cloud-based, which embedding model to choose that could best work with the chosen VD, most importantly the cost factor(free or paid, although as of now most paid ones offer free tiers but scalability needs to be taken into consideration here).

P.S. — To get an understanding of what RAG & Vector database ecosystem is head onto my previous blog posts:

These different features are based upon underlying technologies VDs (Vector Databases) are built. The base is the same for storing high-dimensional space data (such as text, image/video in both structured & unstructured). It allows various semantic search techniques or Nearest Neighbour Search (kNN).

Source: Infosys

Key Parameters:

Open Source: Whether the software is freely available with source code accessible for modification and contribution. Open-source options offer flexibility and customization but may need more professional support.

Hosted Solution: Availability of a managed service by the vendor, eliminating infrastructure management needs. Hosted solutions can be convenient but might restrict control and incur recurring costs.

Pricing: Cost models, including pay-as-you-go options based on usage, tiered subscription plans, or enterprise pricing for custom needs. Consider your expected data volume and usage patterns.

Supported Vector Lengths: The types of vectors (dense with many non-zero values or sparse with mostly zeros) and their maximum lengths supported by the database. This impacts how complex and high-dimensional your data can be.

Supported Distances: Similarity metrics are used for comparing vectors, like Euclidean distance (L2), inner product, or Jaccard similarity. Choose based on the relationships and comparisons relevant to your application.

Nearest Neighbor Search: Methods for finding the closest vectors to a given query vector. Exact methods ensure precision but can be slower, while approximate methods offer faster results with slight accuracy trade-offs.

Clustering: Ability to group similar vectors automatically, useful for data exploration and analysis. Not all vector databases offer built-in clustering functionality.

Filtering & Aggregation: Capabilities to filter data based on specific criteria and perform aggregate operations like calculating average distances. These features can help refine search results and summarize data patterns.

Integrations: Programming languages and SDKs supported for interacting with the database. Choose options compatible with your development environment and tools.

Cloud Providers: Cloud platforms where the database can be deployed. Consider your preferred cloud infrastructure and vendor lock-in potential.

Developer Experience: Ease of use and developer friendliness, including APIs, documentation, and available community support. A smooth development experience can accelerate your project.

Credits : Vector Database Comparison Cheatsheet

References:

--

--

Jayita Bhattacharyya
CodeX
Writer for

Official Code-breaker | Generative AI | Machine Learning | Software Engineer | Traveller