Vector Databases 101What Vectors Are and How They Differ from RDBs and Key–Value Stores

Vector Databases 101 — What Vectors Are and How They Differ from RDBs and Key–Value Stores

Hello everyone. Today, we'll thoroughly learn about "Vector DBs (Vector Databases)" that have become increasingly common to hear about recently. We'll start from the basics: what vectors actually are, then move on to how they differ from RDBs (Relational Databases) and KVS (Key-Value Stores), and finally discuss hardware aspects like "Is CPU alone sufficient? Do we need GPU?" in detail.

The article is quite long (about 5,000 words). Please read it leisurely with a cup of coffee, as if attending a class.

---

1. Introduction: Why "Vector DBs" Now? 💡

In recent years, generative AI and large language models (LLMs) have exploded in popularity. AI systems like ChatGPT and Claude that we interact with daily have learned from vast amounts of text and images, converting them into "numerical values" for processing.

However, AI doesn't directly use "the text itself" or "the image data itself" for calculations. In reality, they convert these into arrays of hundreds to thousands of numerical values (vectors) for processing.

This is where Vector DBs come in. For AI to "find text with similar meaning" or "quickly retrieve related images," we need a database that can efficiently store and search these high-dimensional vectors.

Vector DBs are the infrastructure that enables "semantic search" that traditional RDBs and KVS couldn't handle well.

---

2. What Are Vectors? 🧭

So what exactly are "vectors"?

In mathematics, vectors refer to quantities that have both "magnitude" and "direction." For example, in physics, a velocity like "10 meters per second eastward" is not just a number (scalar quantity) but a vector with direction.

Let me give you a more understandable analogy:

• Scalar: Temperature (30°C), weight (60kg) - things that can be expressed with "just numbers"
• Vector: Wind (10m/s northeast), position (point at x=3, y=4) - things expressed with "collections of numbers" to represent direction or position

In the computer world, we apply this "collection of numbers" more broadly. For example, when AI reads text, it converts that text into high-dimensional vectors (e.g., 768 dimensions, 1024 dimensions).

In other words, a sentence like "I like coffee" is not just a character string but is stored internally in the computer as a vector like [0.12, -0.98, 0.33, …].

---

3. What Is Vector Search? 🔎

Once we can convert text or images into vectors, the next thing we want to do is "find similar things."

Let's consider an example:

• Query: "Photo of a dog running"
• Vectorized images in the database:
• Cat sleeping
• Dog chasing a ball
• Car driving

In this case, what the user is looking for is the "dog chasing a ball" image, right? With traditional search (text matching), it wouldn't hit unless tagged with "dog," but vector search can find semantically similar vectors.

The "closeness" between vectors is measured using mathematical indicators like cosine similarity or Euclidean distance. This allows us to quickly retrieve "semantically similar" items.

Note: The key point is searching by "semantic similarity" rather than exact matches.

---

4. Differences Between Vector DBs and RDBs/KVS 🧩

Now let's get to the main topic. How do Vector DBs differ from traditional databases?

(1) RDB (Relational Database)

• Examples: MySQL, PostgreSQL, Oracle Database
• Stores data in "tables" and manages with rows and columns
• Uses SQL language for operations like "retrieve data that matches WHERE conditions"
• Strengths: Structured data (sales management, customer management, etc.), complex JOIN operations
• Weaknesses: Semantic search. Not good at searching text or images by "similarity"

(2) KVS (Key-Value Store)

• Examples: Redis, Amazon DynamoDB
• Stores data as "key" and "value" pairs
• Example: "user123" → "Name: Yamada Taro, Age: 30"
• Strengths: Ultra-fast simple search (instant value retrieval when you know the key)
• Weaknesses: Complex conditional search, semantic search

(3) Vector DB

• Examples: Pinecone, Milvus, Weaviate, FAISS (library developed by Facebook)
• Stores data as "vectors" and performs fast "similar" searches using Approximate Nearest Neighbor (ANN) search
• Strengths: Semantic search, recommendations (people with similar preferences to you also like this product), image/audio/video search
• Weaknesses: Traditional accounting or inventory management requiring "accurate numerical aggregation"

In summary, RDBs and KVS are "strong with structured data," while Vector DBs are "strong with semantic search."

---

5. Is CPU Alone Sufficient for Vector DBs? ⚙️

Now, here's what might concern you about hardware.

Vector DBs perform large amounts of "similarity calculations between high-dimensional vectors." This is matrix computation, which GPUs excel at.

Can it run on CPU alone?

• Yes, it basically works with CPU alone
• Many Vector DBs like FAISS and Milvus provide CPU version libraries
• For small to medium-scale datasets (hundreds of thousands to millions of items), CPU alone is sufficiently practical

What changes with GPU?

• For large-scale datasets of tens of millions or more, CPU can take time for searches
• GPU can provide parallel computation, making search speeds several to dozens of times faster
• Example: For recommendation services where users submit queries in real-time and get results back, GPU benefits are significant

In actual practice?

• Startups and small-scale development: CPU is sufficient to start. Can scale on cloud
• Large-scale services (hundreds of millions of items): Introduce GPU or distributed clusters

In other words, GPU is not "essential" but an option that "becomes necessary depending on scale."

---

6. Vector DB Use Cases 📚

To deepen our understanding so far, let me introduce several actual use cases:

1. Advanced Search Engines
• Enables "meaning-based search" rather than simple text matching
• Example: Searching "smartphone broken" also hits "mobile repair" and "iPhone repair shop"

1. Recommendation Systems
• "Recommendations for you" on Amazon or Netflix
• Vectorize products or movies and present items similar to user preferences

1. Chatbots and RAG (Retrieval-Augmented Generation)
• Used when giving knowledge bases to AI like ChatGPT
• Search related documents by vector search based on questions and generate answers

1. Image and Audio Search
• Find similar photos, find songs from singing voice
• Can search "similar data" even without text tags

---

7. Summary ✅

Everyone, I think "what Vector DBs are" has become much clearer by now. Let me organize the key points:

1. What are vectors?

Collections of numbers; AI converts text and images into vectors for processing

1. Role of Vector DBs

Databases for finding semantically "similar" things

1. Differences from RDBs and KVS
• RDB: Table format, strong with structured data aggregation
• KVS: Strong with ultra-fast search by key matching
• Vector DB: Strong with semantic search

1. CPU vs GPU
• Small scale: CPU is sufficient
• Large scale/real-time processing: GPU is effective

1. Use Cases

Search engines, recommendations, chatbots, image search, etc. - rapidly expanding across all AI fields

Vector DBs can be said to be the "new foundation for search in the AI era." I encourage you to try them in small projects while being aware of their differences from RDBs and KVS - this will deepen your understanding even further.