Vector Databases A Practical Introduction (Nitin Borwankar)（Z-Library） (1)

Vector Databases A Practical Introduction Nitin Borwankar

Vector Databases by Nitin Borwankar Copyright © 2026 Nitin Borwankar. All rights reserved. Published by O’Reilly Media, Inc., 141 Stony Circle, Suite 195, Santa Rosa, CA 95401. O’Reilly books may be purchased for educational, business, or sales promotional use. Online editions are also available for most titles (http://oreilly.com). For more information, contact our corporate/institutional sales department: 800-998-9938 or corporate@oreilly.com. Acquisitions Editor: Andy Kwan Development Editor: Sara Hunter Production Editor: Beth Kelly Copyeditor: Audrey Doyle Proofreader: Krsta Technology Solutions Indexer: WordCo Indexing Services, Inc. Cover Designer: Susan Thompson Cover Illustrator: José Marzan Jr. Interior Designer: David Futato Interior Illustrator: Kate Dullea April 2026: First Edition Revision History for the First Edition

2026-04-06: First Release See http://oreilly.com/catalog/errata.csp?isbn=9781098177591 for release details. The O’Reilly logo is a registered trademark of O’Reilly Media, Inc. Vector Databases, the cover image, and related trade dress are trademarks of O’Reilly Media, Inc. The views expressed in this work are those of the author and do not represent the publisher’s views. While the publisher and the author have used good faith efforts to ensure that the information and instructions contained in this work are accurate, the publisher and the author disclaim all responsibility for errors or omissions, including without limitation responsibility for damages resulting from the use of or reliance on this work. Use of the information and instructions contained in this work is at your own risk. If any code samples or other technology this work contains or describes is subject to open source licenses or the intellectual property rights of others, it is your responsibility to ensure that your use thereof complies with such licenses and/or rights. 978-1-098-17759-1 [LSI]

Preface To repeat a well-worn phrase, this is the book I wish I had when I started exploring vector databases. I was looking for a single source of basic theory and practical applications of vector databases, but I found none. Existing vector database resources are fragmented across vendor documentation, blog posts, YouTube videos, and GitHub code. I wanted practical working examples rather than theoretical coverage. I also have a personal preference for proceeding systematically from basics to simple working systems to help organize data locally—especially private data. Thus began my journey to pull together information from multiple sources and multiple directions into a coherent whole. And this book is what emerged. It is meant to be useful to the intermediate developer, from the theoretically minded to the “just show me the code” type. There’s something for everyone. The book follows a clear pedagogical arc: concepts → tools → applications → future directions. Each chapter builds on previous ones while remaining practical with working code examples. The progression moves from understanding vector databases to building working applications that are meant for personal use, hands-on interaction, and extension by developers who want to build personal data management systems. It culminates in a forward-looking proposal for industry standardization of a Vector Query Language modeled after SQL. The somewhat opinionated part of this book’s approach is a focus on simple hybrid vector relational databases—created by adding vector extension plug-ins to the well-known open source databases SQLite3 and PostgreSQL. The reason for this is that these databases are well known among database developers. Simply adding a data type of “vector” opens up a whole new world of AI data applications. This makes onboarding as low- effort as possible.

NOTE It can’t be emphasized enough that these applications are for hands-on learning and not for web-scale production deployment. Modify them as you see fit; make them your own. And now let’s dive into the contents. What’s in This Book Chapters 1 and 2 establish foundational concepts—vector databases, embeddings, and semantic search fundamentals: Chapter 1, “Introduction to Vector Databases” Introduces vector databases as foundational technology for AI applications, explaining why unstructured data (text, images, audio, video) requires semantic search rather than keyword matching Covers the vector data type, similarity search, and key operations (cosine similarity, nearest-neighbor search, vector arithmetic) Contrasts vector databases with SQL and NoSQL databases, highlighting hybrid architectures that combine structured metadata with vector similarity search Establishes the conceptual foundation for understanding how embeddings preserve semantic meaning in vector space Chapter 2, “Embeddings” Takes a deep dive into the history, evolution, and current usage of vector embeddings with code examples for language model applications

Explains how embeddings bridge the gap between unstructured data and machine-processable representations, mapping raw data into vectors that capture semantic relationships Covers the mathematical foundations and how proximity in vector space reflects semantic similarity Provides the conceptual framework needed for practical applications later in the book Chapters 3 and 4 discuss practical implementations with Facebook AI Similarity Search (FAISS) and SQLite, showing how to build personal-scale search systems: Chapter 3, “Similarity Search with FAISS” Explores FAISS as both a production similarity search engine and a flexible toolkit for building custom vector databases Is designed for developers interested in the internals of vector search and those wanting to extend capabilities of systems such as sqlite- vss Covers the mathematical foundations and implementation details of efficient similarity search at scale Focuses on CPU-based implementations for maximum accessibility Chapter 4, “Semantic Search with SQLite3” Builds a personal knowledge management system using semantic search over Reddit content, demonstrating meaning-based search versus keyword matching

Introduces sqlite-vss, a SQLite extension that wraps FAISS and integrates vector search into the relational database world Combines semantic retrieval with relational metadata filtering in a single SQL workflow Uses an “overfetch-then-filter” pattern to handle the fact that metadata filters aren’t pushed down into FAISS search Chapters 5 through 8 progress through increasingly sophisticated applications—ArXiv paper search, local retrieval-augmented generation (RAG) systems, scientific RAG, and personal conversation search—moving from public to private data: Chapter 5, “Building an ArXiv Paper Search System with PostgreSQL pgvector” Constructs a scientific literature search system addressing the challenge of discovering relevant research across millions of papers Demonstrates how vector embeddings capture semantic relationships that keyword search misses (e.g., connecting “neural network optimization” with “gradient descent improvements”) Leverages PostgreSQL with the pgvector extension for enterprise- scale applications Shows how to handle technical terminology, structured academic content, and metadata in a unified search system Chapter 6, “Building a Retrieval-Augmented Generation System with SQLite VSS and Ollama”

Assembles a complete RAG system combining vector search with local large language model (LLM) capabilities Builds an entirely local, private system running on a single desktop using SQLite Vector Similarity Search (SQLite VSS) and Ollama Addresses the fundamental LLM limitation: knowledge frozen at training time without access to private or recent information Creates a question-answering system that grounds LLM responses in actual retrieved data to reduce hallucinations Chapter 7, “Building a Scientific RAG System with PostgreSQL and pgvector” Develops a RAG system specifically designed for scientific literature with unique challenges: technical terminology, structured content, citation networks, and evidence quality Enables semantic discovery, cross-paper synthesis, and contextual understanding across ArXiv’s 15,000+ monthly publications Uses PostgreSQL with pgvector for scientific knowledge retrieval Demonstrates handling domain-specific language and structured academic conventions Chapter 8, “Building a Complete Conversation Search and RAG System” Moves from searching public data to building a “second brain” for personal chat history with AI assistants

Addresses unique characteristics of conversational data: contextual dependencies, conversational flow across exchanges, personal language patterns, and privacy requirements Handles the messiness of real conversational data where ideas develop across multiple message exchanges Creates a system for retrieving valuable insights, solutions, and learned knowledge from thousands of personal AI conversations Chapter 9 proposes an experimental Vector Query Language (VQL) to standardize the fragmented vector database landscape: Chapter 9, “Vector Query Language” Presents a prototype SQL-inspired query language for vector databases, designed to start community discussion and consensus building Addresses the needs of three communities: application developers seeking database abstractions, database maintainers needing cross- vendor tooling, and AI/machine learning researchers requiring efficient data manipulation Defines a data model (collections, tables, vectors, embeddings, metadata) and query syntax combining SQL familiarity with vector operations Covers similarity search, hybrid search, range search, batch operations, vector functions, and aggregations Proposes a path toward standardization across the fragmented vector database landscape

Who This Book Is For This book’s target audience comprises: Enterprise software developers building their first AI applications Data scientists implementing semantic search Machine learning engineers working on RAG systems Backend developers adding vector capabilities to existing systems Additionally, it is geared toward people who learn by doing, so there are also practical applications focused on individual use. Prerequisites include Python programming proficiency, a basic understanding of databases and SQL, and familiarity with running LLMs locally for the RAG chapters. Readers should also be comfortable with simple mathematical concepts. The book is for those who want to learn how to build small- to medium- scale semantic search systems, implement RAG applications with local LLMs, optimize vector similarity search, choose the right indexes, and integrate vector search with relational databases. The book focuses on Python implementations rather than other languages, covers CPU-based solutions rather than GPU optimization, and excludes cloud-managed vector database services in favor of self-hosted solutions. The scope of the applications and the theory is to cover enough to get started and do something useful with personal data as a first step. This is not a textbook on how to build web-scale production vector database applications. However, the lessons in the chapters can serve as a foundation and a springboard to take the next leap up in scope and complexity. How to Use This Book For the reader who is coming from a purely relational database background, reading the first two chapters is a must. Chapter 3 is more detailed on the

innards of FAISS, but you can just use the code examples to dive right in and use it. Chapters 4 through 8 use SQLite3 and Postgres to build semantic search and RAG applications using local data and local LLMs. Chapter 8 is especially topical, as it creates a semantic search over your LLM conversations—in this case, exported from Claude. Finally, those interested in exploring similarities with SQL can dive into Chapter 9, which is a speculative design of a Vector Query Language to enable a unified, vendor-independent, client access metaphor for vector databases. Software, Environment, and Resource Requirements With the following resources, the example applications will run on standard developer laptops: Python version 3.11+ Key dependencies: PostgreSQL, SQLite3, FAISS, Ollama OS: Linux, macOS, or WSL on Windows Minimum 16 GB RAM; preferably 24 GB For a developer machine that can run PostgreSQL and an 8B LLM: Storage requirements: 512 GB to 1 TB Conventions Used in This Book The following typographical conventions are used in this book: Italic Indicates new terms, URLs, email addresses, filenames, and file extensions.

Constant width Used for program listings, as well as within paragraphs to refer to program elements such as variable or function names, databases, data types, environment variables, statements, and keywords. Constant width bold Shows commands or other text that should be typed literally by the user. Constant width italic Shows text that should be replaced with user-supplied values or by values determined by context. TIP This element signifies a tip or suggestion. NOTE This element signifies a general note. WARNING This element indicates a warning or caution. Using Code Examples Supplemental material (code examples, exercises, etc.) is available for download at https://github.com/nborwankar/VectorDatabaseBook. If you have a technical question or a problem using the code examples, please send email to support@oreilly.com.

This book is here to help you get your job done. In general, if example code is offered with this book, you may use it in your programs and documentation. You do not need to contact us for permission unless you’re reproducing a significant portion of the code. For example, writing a program that uses several chunks of code from this book does not require permission. Selling or distributing examples from O’Reilly books does require permission. Answering a question by citing this book and quoting example code does not require permission. Incorporating a significant amount of example code from this book into your product’s documentation does require permission. We appreciate, but generally do not require, attribution. An attribution usually includes the title, author, publisher, and ISBN. For example: “Vector Databases by Nitin Borwankar (O’Reilly). Copyright 2026 Nitin Borwankar, 978-1-098-17759-1.” If you feel your use of code examples falls outside fair use or the permission given above, feel free to contact us at permissions@oreilly.com. O’Reilly Online Learning NOTE For more than 40 years, O’Reilly Media has provided technology and business training, knowledge, and insight to help companies succeed. Our unique network of experts and innovators share their knowledge and expertise through books, articles, and our online learning platform. O’Reilly’s online learning platform gives you on-demand access to live training courses, in-depth learning paths, interactive coding environments, and a vast collection of text and video from O’Reilly and 200+ other publishers. For more information, visit https://oreilly.com.

How to Contact Us Please address comments and questions concerning this book to the publisher: O’Reilly Media, Inc. 141 Stony Circle, Suite 195 Santa Rosa, CA 95401 800-889-8969 (in the United States or Canada) 707-827-7019 (international or local) 707-829-0104 (fax) support@oreilly.com https://oreilly.com/about/contact.html We have a web page for this book, where we list errata and any additional information. You can access this page at https://oreil.ly/vector-databases. For news and information about our books and courses, visit https://oreilly.com. Find us on LinkedIn: https://linkedin.com/company/oreilly-media Watch us on YouTube: https://youtube.com/oreillymedia Acknowledgments I want to thank all the technical reviewers, especially Sunil Sawant. Because of their efforts, the book is a much better product.

Sara Hunter, my development editor, deserves praise for her patience during the yo-yo delivery cadence of the chapters. I would also like to thank the rest of the O’Reilly editorial team—both the early release and final production teams—for putting up with the inexperience of a first-time author. My wife, Garima, and daughter, Vanita, had to tolerate my minimal participation in family events during all of last year. I would like to thank them deeply for their support and understanding during the writing of “the book.” It’s finally done, guys! Of course, none of this would be possible without the open source developer communities of PostgreSQL, SQLite, and the vector extensions, whose work enabled the example code.

Chapter 1. Introduction to Vector Databases The current wave of generative AI (GenAI for short), especially in natural language applications, has brought the technology of vector databases to the forefront of developer attention. Unstructured data used to train models is vectorized, and data used in inference is often saved in vector databases. And yet the technology is not well understood or widely known to the average database application developer. This chapter provides a high-level overview of vector databases, outlining their main components and offering a preview of the material ahead. Mostly, you’ll learn what the technology is and why you need it. By the end of the chapter, you’ll understand how to use vector databases; the key topic of embeddings will be covered in the next chapter. Why Do You Need Vector Databases? Even before the advent of GenAI, managing unstructured data such as text, images, audio, and video was difficult. It required specialized technology and knowledge to manage these assets at scale. Indexing and searching over such data were only possible at the level of the metadata—for example, the date of creation, size, author, and so on. But searching for an image similar to another one, or a paragraph with a similar meaning to another, was not something an SQL database could do. SQL databases are used to match on syntax, not semantics. Now, with the emergence of vector databases into the mainstream, managing unstructured data and searching based on semantics—the meaning of the data—have become much easier. This is the world of vector databases, which opens up new possibilities for applications based on semantic or similarity search. This goes well beyond searching by keyword,

which often leads to confusion between, say, apple as a fruit and Apple as a computer company. Vector databases are a foundational technology for AI in the enterprise, just as relational databases are the foundational technology for business data. Applications are now being built that combine business data with the semantics of text, images, audio, and video to create rich, intelligent, interactive systems. To make these rich applications, you’ll need a deeper understanding of vector databases than is broadly available today. While specialized vector databases are currently available (circa 2025), they all differ slightly and require database-specific knowledge. Many such vector databases are available in the cloud—requiring far less specialized maintenance—while databases such as Postgres, enhanced with vector capability, are similarly available. This book focuses on technology in which vector database capability is well integrated into the world of relational databases. This allows structured and unstructured data to be handled in familiar ways and managed using familiar tools for databases, such as backup, restore, partitioning, and access control, among others. Vector (i.e., unstructured) data is saved in familiar database tables without special treatment based on what these tables might contain. Prior to vector databases, unstructured data was often stored as BLOBs (binary large objects) or directly in files, with a field in a relational database pointing to the file. In either case, the data was considered opaque—hence a BLOB. We were not able to do much with the data aside from saving and retrieving it, except perhaps with text, where keyword search was possible. A New Data Type: Vector So what is “vector data,” anyway, and why can’t we just use the BLOB as it is? We could—but with a vector, we get a superpower that allows us to add a semantic component to our data and also make it GenAI-ready. In the language model part of GenAI, words, sentences, paragraphs, and even documents may be represented as vectors—so to do the heavy lifting in the

GenAI space, it’s essential to understand vector data. Let’s dive a little deeper. First, to start with a familiar domain, think of “vector” as a new data type in a database management system. As a concrete example, let’s consider Postgres, which will be our mainstay for this book and is a well-known and popular database. In Postgres, we can add new types to the type system via the Postgres extension system. In the case of the vector type, the pgvector extension provides the capability to create columns of type VECTOR in a table and offers a vector-native search capability—similarity search—over this data type. If you look at the content of a VECTOR column, as in Figure 1-1, you will see a list of floats—but there’s more to this than meets the eye. Figure 1-1. Vector data type contents Similarity Search The similarity search capability is what differentiates the VECTOR type from merely saving and fetching an ordered set of floats. It is this similarity search that allows you to find text similar in meaning to a given piece of text. This vector similarity capability forms the foundation of database

applications built with large language models (LLMs) such as ChatGPT and Claude, or local LLMs like Llama3.x. The vector similarity search has its own search operator, which the pgvector extension adds to the type system of Postgres. We will explore this in more detail in the next chapter on vector embeddings. Similarity search example Imagine you are searching a company’s knowledge base, and you type, “How do I get my money back?” See Figure 1-2 for what happens next. Figure 1-2. Embedding process flow Keyword search (literal word matching) In a keyword search, the system looks for the exact words you typed (or close variants). It performs well if a page literally contains words such as “money” and “back” or the phrase “get my money back.” However, it may miss the best answer if the document uses different wording, such as: “Refund policy” “Return and exchange” “Reimbursement”