Vector Databases: Introduction

Once you've turned data into embeddings, you need somewhere to store and search them — fast. A vector database is built for exactly this: it stores embedding vectors and finds the ones most similar to a query, powering semantic search, RAG, and recommendations at scale. It's the storage layer that makes embeddings useful in real applications.

💡 In one line: A vector database stores embedding vectors and searches them by similarity — finding the nearest matches by meaning, not exact text.

What is a Vector Database?

It's a database designed to store embedding vectors and search them by similarity (nearest neighbours) rather than exact match. Each record holds a vector, an ID, and metadata, and the whole store is indexed so similarity search stays fast even across millions of vectors.

Why Not a Regular Database?

Traditional databases are great at exact and keyword lookups (WHERE title = 'cat'). But they can't efficiently answer "find the most similar by meaning" across millions of high-dimensional vectors. Vector databases use special approximate nearest-neighbour (ANN) indexes to do this quickly.

Remove

What It Stores

Each record typically has three parts:

• ID — a unique key.
• Vector — the embedding.
• Metadata — extra fields like the original text, source, tags, or timestamps.

Records are grouped into collections (or indexes / namespaces) — more on that later.

How It Works

The core flow stores data as vectors, then searches by embedding the query. 

Whiteboard

Key Operations

• Upsert — insert or update vectors (+ metadata).
• Query / Search — find the nearest neighbours to a query vector.
• Filter — restrict results by metadata.
• Delete — remove records.

Vector DB vs. Traditional DB

Popular Vector Databases

Common options include Pinecone, Weaviate, Qdrant, Milvus, Chroma, the pgvector extension for PostgreSQL, and the FAISS library. They range from managed cloud services to self-hosted and embedded libraries.

Use Cases

• RAG — retrieve relevant context for an LLM.
• Semantic search — find by meaning, not keywords.
• Recommendations, deduplication, image/audio search, and anomaly detection.

What's Ahead in This Topic

• Storing embeddings and metadata storage.
• Collections & namespaces for organising data.
• Similarity search, k-nearest neighbours, top-k retrieval, and ANN search.

Summary

• A vector database stores embeddings and searches them by similarity.
• Unlike traditional DBs (exact match), it finds the nearest vectors by meaning using ANN indexes.
• Each record holds an ID, vector, and metadata, grouped into collections.
• Core operations: upsert, query, filter, and delete.
• It's the backbone of RAG, semantic search, and recommendations. EOF echo created