Vector

Databases

Dr. Navneet Goyal

Senior Professor & Head
Department of Computer Science
BITS Pilani, Pilani Campus
 Evolution of Database Technology

Image source: [Link]

5/5/2026 SAMPLE FOOTER TEXT 2

 Vector Databases
• What are Vector Databases?
• What Are Embeddings?
• How do vector databases work?
• Use cases of Vector Databases
• Open problems

SAMPLE FOOTER TEXT 3

5/5/2026
 Vector Databases: Motivation
• Which two clients are similar in terms of their requirements?
+ More generally, which clients have similar requirements
• Which two employees have similar skills & experience?
• Suppose an employee decides to leave HCL, and you need to find a replacement!
• Suppose a software which you want to buy is very expensive! You want to replace it with its “closest”
cheaper version!
• Recommender Systems – which product(s)/service (s) similar to one purchased by a customer would
you recommend?
• Image-based retrieval
• Find documents that are similar to a given document based on their topicand sentiment

4
5/5/2026
 Vector Databases
• A vector database is a fully managed, no-frills solution for storing, indexing and
searching across a massive dataset of unstructured data that leverages the power
of embeddings from machine learning models
• A type of database that stores data as high-dimensional vectors, which are
mathematical representations of features or attributes. Each vector has a certain
number of dimensions, which can range from tens to thousands, depending on the
complexity and granularity of the data
• The vectors are usually generated by applying some kind of transformation or
embedding function to the raw data, such as text, images, audio, video, and others.
The embedding function can be based on various methods, such as machine/deep
learning models, word embeddings, feature extraction algorithms

SAMPLE FOOTER TEXT 5

5/5/2026
 Vector Databases vs. Traditional
Databases

• Traditional relational database systems excel at managing

structured data with predefined formats and executing precise
search operations.
• In contrast, vector databases specialize in storing and retrieving
unstructured data types, such as images, audio, videos, and
textual content, through high-dimensional numerical
representations known as vector embeddings.

SAMPLE FOOTER TEXT 6

5/5/2026
 Vector Databases: Operations
• Vector databases are used to perform semantic similarity searches
using techniques like the Approximate Nearest Neighbor
(ANN) algorithm.
• Applications in various domains:
• recommender systems
• chatbots
• searching for similar images, videos, and audio content.
• With the rise of AI and large language models (LLMs) like ChatGPT,
vector databases are also beneficial in addressing LLM hallucinations.

SAMPLE FOOTER TEXT 7

5/5/2026
 Approximate Nearest Neighbour (VDBMSs)
& Approximate Queries (RDBMS)
• Vector databases are used to perform semantic similarity searches using
techniques like the Approximate Nearest Neighbor (ANN) algorithm.
• Applications in various domains
• recommender systems
• chatbots
• searching for similar images, videos, and audio content.
• With the rise of AI and large language models (LLMs) like ChatGPT, vector
databases are also beneficial in addressing LLM hallucinations.

SAMPLE FOOTER TEXT 8

5/5/2026
Approximate vs. Exact Queries (RDBMS)
SQL Query
Exact Answer
Long Response Times!

• Exact answers NOT always required

+ DSS applications usually exploratory: early feedback to help identify “interesting”
regions
+ Aggregate queries: precision to “last decimal” not needed
• e.g., “What percentage of the US sales are in NJ?” (display as bar graph)
+ Preview answers while waiting. Trial queries
+ Base data can be remote or unavailable: approximate processing using locally-cached
data synopses is the only option
Fast Approximate Answers
• Primarily for Aggregate queries
• Goal is to quickly report the leading digits of answers
+ In seconds instead of minutes or hours
+ Most useful if can provide error guarantees
E.g., Average salary
$59,000 +/- $500 (with 95% confidence) in 10 seconds
vs. $59,152.25 in 10 minutes
• Achieved by answering the query based on samples or other synopses of the
data
• Speed-up obtained because synopses are orders of magnitude smaller than
the original data
• Online Aggregate Queries
Approximate NN Search
• ANN aims to find an approximate nearest neighbour(s) with high
probability while minimizing the computational cost
• Same objective as that for Approximate Queries
• NN search is exhaustive and computationally infeasible for large
datasets
• Can be optimized using spatial indexing structures like R-Trees,
kD-Trees, Grid-R-Tree (our work), DDR-Tree (our work) etc.
 Vector Databases

Image source: [Link]

5/5/2026 SAMPLE FOOTER TEXT 12

 Desirable Features of a VDBMS
1. Scalability
2. Vector similarity Search (k-NN, ANN)
3. The need to support not only fast query processing on large-scale
vector data but also the efficient handling of dynamic vector data
(such as insertions and deletions).
• Youtube uploads 500 hours of user-generated videos per minute and meanwhile
offers real-time recommendations.

SAMPLE FOOTER TEXT 13

5/5/2026
 Desirable Features of a VDBMS
4. The need to provide advanced query processing such as attribute
filtering and multi-vector query processing
• attribute filtering is to only search vectors that satisfy a given filtering condition, which is useful in e-
commerce applications e.g., finding the T-shirts similar to a given image vector that also cost less
than $100
• multi-vector query processing targets for the scenario where each object is described by multiple
vectors, e.g., profiling a person using a face vector and a posture vector in many computer vision
applications

SAMPLE FOOTER TEXT 14

5/5/2026
 Desirable Features of a VDBMS
5. Support for Hybrid Data (Vector, non-vector, & Indexes)
• Each entity in Milvus* is described as one or more vectors and optionally some numerical
attributes
• For example, in the image search application, the numerical attributes can represent the age
and height of a person in addition to possibly multiple machine-learned feature vectors of
his/her photos (e.g., describing front-face, side-face, or posture)
• In future – categorical attributes along with indexes like inverted lists or bitmaps

*Milvus: A Purpose-Built Vector Data Management System, Jianguo Wang et al. SIGMOD 2021

SAMPLE FOOTER TEXT 15

5/5/2026
 What are Embeddings?
2 Main trends:
• According to IDC, 80% of data will be unstructured by 2025*
• Surge in unstructured data demands a paradigm shift in DB Technology: VECTOR DATABASES
• Unprecedented developments in Machine Learning which can transform unstructured data into
learned feature vectors

• [Link]
SAMPLE FOOTER TEXT 16
5/5/2026
 What are Embeddings?
• An embedding is a form of data transformation where high-dimensional data is mapped
onto a lower-dimensional vector without losing much of its original information.
• Examples: Item2vec, word2vec , doc2vec, graph2vec etc.
• YouTube embeds videos to vectors
• Airbnb models houses with vectors
• Bioscientists describe the molecular structural information of drug compounds
using vectors
• Images and texts are also naturally represented by vectors

SAMPLE FOOTER TEXT 17

5/5/2026
 What are Embeddings?
• Word2Vec - Encodes words based on surrounding context patterns. Words with
similar meanings have similar vectors.
• Doc2Vec - Extends Word2Vec by embedding entire documents. Documents about
similar topics cluster together.
• Image Embeddings - Encode images based on visual features like objects,
scenes, and textures detected by a neural net.
• Graph Embeddings - Represent nodes in a graph based on connection patterns
and node attributes
• Multimodal data Embeddings – How to represent multimodal data as vectors?

SAMPLE FOOTER TEXT 18

5/5/2026
 What are Embeddings?

Embeddings convert messy real-world data into mathematical representations

capturing hidden relationships. This transformed data powers cutting-edge AI
• How to store, manage, retrieve, “query”, or run algorithms on these vectors?
• Vector Databases (VDBMS)

SAMPLE FOOTER TEXT 19

5/5/2026
 Our Recent work in Databases

• Data Provenance Models for SQL & NoSQL Databases

• Zero Information Loss Databases
• One PhD thesis: Design And Development Of Data Provenance Frameworks For SQL And
NoSQL Databases Using Zero-informationLoss Database Concept

• Indexing structures for speeding up ML algorithms

• One PhD thesis: Design and Development of Data Indexing Techniques for Mining Large
and Streaming Data
• Grid-R-Tree & DDR-Tree

SAMPLE FOOTER TEXT 20

5/5/2026
 Our Recent work in Learning Embeddings
• High-resolution Satellite Image Time series (AAAI 2024)
• Learning embeddings for Multimodal Data

SAMPLE FOOTER TEXT 21

5/5/2026
 Future of AI/Databases
• Vector AI
• Marriage of of AI & DB Technologies - AI4DB & DB4AI
• YeSQL: User-Defined Functions in Relational Databases: Challenges and Promising
Solutions based on YeSQL: Keynote talk by Yannis Ioannidis (ACM President) @ IEEE
DSAA 2023

SAMPLE FOOTER TEXT 22

5/5/2026
5/5/2026

Concluding Remarks
• As AI adoption grows, purpose-built vector data platforms will become crucial
components of enterprise data infrastructure.
• Combining the right database technologies with advanced embedding
techniques will unleash deeper insights and performance gains.
• The companies that learn to store, organize, and query vector data most
effectively will gain a competitive edge.
• Vector databases and embeddings provide the data foundations for an AI-
first future

[Link] newsletter
SAMPLE FOOTER TEXT 23
5/5/2026

Popular VDBMSs
• Pinecone, Chroma, Milvus, Weaviate, Vald, Qdrant and Deep Lake.
However,
• Vald – more of a vector search engine!
+ primarily focuses on similarity search and lacks features such as access control
and integrations to other technologies

SAMPLE FOOTER TEXT 24

5/5/2026

Use cases of VDBMSs

• Similarity search
• Image and video similarity search
• Voice recognition
• Chatbots and long-term memory
• Enable large language models (LLMs) to generate more relevant and
coherent text based on an AI plugin.

SAMPLE FOOTER TEXT 25

5/5/2026

Image Search

SAMPLE FOOTER TEXT 26

5/5/2026

Current Challenges with VDBMSs

• Balancing between speed and accuracy
• Growing dimensionality and sparsity
• VDBMS vendors need to focus on developing and applying new algorithms for
vector indices, making high-dimensional vectors more human-readable through
visualizations.
• Additionally, as data-intensive computational models are computationally expensive
to retrain, we expect that VDBMS vendors will try to address this by implementing
features for incremental learning, i.e., cost-effective fine-tuning of computational
models

SAMPLE FOOTER TEXT 27