Vector Databases

Introduction

A vector database stores embeddings and finds nearest neighbours fast — the core operation behind RAG, semantic search, and recommendation. Modern options range from managed (Pinecone) to self-hosted (Qdrant, Chroma) to Postgres extensions (pgvector).

Beginner analogy: Like a search engine, but it indexes meaning instead of keywords.

Understanding the topic

Core concepts to understand:

• Pinecone — managed, fast, easy.
• Chroma — open-source, great for prototyping.
• Qdrant, Weaviate, Milvus — open-source, production-grade.
• pgvector — Postgres extension, lives in your existing DB.
• Lovable Cloud (Supabase) ships pgvector — perfect for app DB + vectors in one.

Syntax reference

Visual workflow / architecture:

   Vectors                 Index
   ───────                ───────
   v1 [0.1,...]    →     HNSW / IVF
   v2 [0.3,...]    →     k-NN search
   v3 [0.9,...]
       ...
                          │
                  query ──┘
                          │
                          ▼
                  top-k similar vectors

Real-world use

Notion AI reportedly uses pgvector. Perplexity uses Pinecone. Vercel docs Q&A uses Postgres + pgvector. OpenAI's Assistants API abstracts the vector DB.

Best practices

• Start with pgvector if you already have Postgres — fewer moving parts.
• Pick HNSW index for speed; IVF for memory.
• Always benchmark recall@k on your real queries.

Hands-on exercise

Interview preparation — practice these questions:

• Q1. What does a vector database do?
• Q2. Compare pgvector vs Pinecone.
• Q3. What is HNSW?