VectorDB (Vector Database)

kc@example.com (kc kim) — Thu, 06 Nov 2025 18:12:00 +0900

VectorDB (vector database) is a database that has recently become very important in AI, machine learning, and LLM (large language model) fields.
Simply put, it is a database specialized for efficiently storing data in numeric vector form, or embeddings, and quickly finding similar data.

What Is VectorDB?

VectorDB is a DB optimized for storing, searching, and comparing vector data.
Usually, unstructured data such as text, images, audio, and code is converted into fixed-length vectors, or numeric arrays, through an embedding model, and these vectors are stored in VectorDB.

For example:

"How is the weather today?"  ->  [0.12, -0.08, 0.56, ... , 0.33]
"Is it raining now?"         ->  [0.11, -0.07, 0.57, ... , 0.31]

These two vectors are semantically similar, so the distance between the vectors is very close.
VectorDB is a system designed to perform this kind of “similarity-based search” quickly.

Key Features of VectorDB

Feature	Description
Embedding storage	Stores vectors generated from text, images, and other data
Similarity Search	Quickly finds the vectors, or data, most similar to a query vector
Metadata filtering	Stores additional information such as tags and categories together with vectors and enables filtering during search
Clustering/classification	Groups semantically similar vectors or classifies them by category
LLM integration	Often used to implement LLM “context” or “memory” features, such as RAG architectures

Why VectorDB Matters

Large language models such as GPT, Claude, and Gemini have limits on the amount of input they can process at once, known as context length.
Therefore, you cannot put all information into an LLM.

In this case, VectorDB is used as follows:

User question -> converted to embedding -> related documents searched in VectorDB ->
results passed to the LLM together with the question to generate an answer. This structure is called RAG, or Retrieval-Augmented Generation.

In other words, VectorDB works like external memory for an LLM.

Representative VectorDB Products

Product	Features
Pinecone	Fully managed, API-centric, specialized for RAG
Weaviate	Open source, supports GraphQL API, supports hybrid search
Milvus	Strong for large-scale data processing, open source
Qdrant	Rust-based, high performance, easy REST/gRPC API
FAISS	Library created by Meta, closer to a search engine than a DB
Chroma	Python-friendly, integrates easily with LangChain and similar tools
Redis Vector Search (Redis 7.2+)	Provides vector indexing inside Redis

Core Concepts in VectorDB Search

Vector Similarity Calculation Methods

Cosine similarity: How similar the directions of two vectors are
Euclidean distance: The straight-line distance between two vectors
Dot product: Mainly used with trained embeddings

Indexing

Because vectors are usually stored in the thousands to millions, ANN (Approximate Nearest Neighbor) algorithms are used to make search fast.

Representative index algorithms:

HNSW (Hierarchical Navigable Small World)
IVF (Inverted File Index)
PQ (Product Quantization)

VectorDB + LLM Example (RAG Architecture)

flowchart LR
A[User question] --> B[Embedding generator]
B --> C[Search similar documents in VectorDB]
C --> D[Return related documents]
D --> E[Send question with documents to LLM]
E --> F[Generate accurate and contextual answer]

With this structure, you can build a knowledge-based chatbot similar to ChatGPT yourself.

Practical Example (Python + Qdrant)

from qdrant_client import QdrantClient
from qdrant_client.models import PointStruct

client = QdrantClient(":memory:")

# Store example vectors
client.upsert(
    collection_name="docs",
    points=[
        PointStruct(id=1, vector=[0.1, 0.3, 0.5], payload={"text": "Hello"}),
        PointStruct(id=2, vector=[0.11, 0.29, 0.51], payload={"text": "Nice to meet you"})
    ]
)

# Search for data similar to the query vector
hits = client.search(collection_name="docs", query_vector=[0.12, 0.28, 0.49], limit=2)
print(hits)

Summary

Item	Description
Purpose	Vector-based similarity search and AI applications
Data format	High-dimensional vectors (embeddings)
Main use cases	RAG, recommendation systems, image search, voice search
Core technologies	ANN indexing, cosine similarity, metadata filters
Representative products	Pinecone, Qdrant, Milvus, Weaviate, Chroma

devkuma – VectorDB