🤖 AI / ML

Vector Database

Vector Store · Vector Search Engine

Definition

A vector database is specialised storage optimised for indexing and nearest-neighbour search in high-dimensional vector spaces at sub-second latency.

What a vector database is

A vector database stores objects as numeric vectors — embeddings — and specialises in one job:
finding the most similar objects to a query vector, quickly.

The standard flow in e-commerce:

Product → embedding model → vector [0.23, -0.41, 0.87, ...] → vector database
Query   → embedding model → vector [0.21, -0.39, 0.85, ...] → ANN search → top 20 products

Cosine distance between vectors is the measure of semantic closeness. Two vectors near each other in
the space mean two objects similar in meaning, regardless of how they are worded.

Why a specialised database is needed

A relational database is a poor fit for similarity search in high-dimensional spaces. At a million
products and 768-dimensional vectors, an exact scan on every query would take seconds. Vector
databases use dedicated indexes:

Index	Principle	When it fits
HNSW	A graph of nearest neighbours	High accuracy, moderate volume
IVF	Clustering plus in-cluster search	Very large catalogues
PQ	Vector quantisation	Memory savings

HNSW (hierarchical navigable small world) is the most widely used: under 10 ms lookup at recall
above 95%.

Applications in e-commerce

Semantic search. A query such as a warm jacket for mountain hiking becomes an embedding and is
matched against the catalogue of product descriptions. Relevant results appear even with no keyword
overlap.

Similar items recommendations. The embedding of the viewed item drives a nearest-neighbour
lookup in product space. It works on cold items with no interaction history.

AI shopping assistants and RAG. An LLM assistant receives context through retrieval — a vector
search across descriptions, specifications and product FAQs.

Tip: for hybrid search (vector relevance plus keywords plus filters), choose a database with
native hybrid support, or pair pgvector with PostgreSQL full-text search.

Common implementations

Qdrant — Rust, high performance, open source, a good fit for self-hosted deployment
Pinecone — a managed service, simple integration, serverless tier
Weaviate — hybrid search, modular architecture
Milvus — scales to very large collections
pgvector — a PostgreSQL extension, convenient when the infrastructure is already on Postgres

Frequently asked questions

How does a vector database differ from a relational one?

A relational database searches exact values or ranges and scales poorly to similarity search in high-dimensional spaces. A vector database stores numeric representations of objects and is optimised for approximate nearest neighbour search — finding the most similar items in milliseconds. For the task find 20 products similar to this one, a vector database is the only practical choice on a large catalogue.

How is a vector database used in semantic search?

The query and the product descriptions pass through the same embedding model and become vectors. The database finds the products with the smallest cosine distance to the query vector. That returns relevant results even when the keywords do not match — comfortable shoes for long walks finds trail running shoes.

Which vector databases are common in e-commerce?

Qdrant (Rust, high performance, open source), Pinecone (managed, simple to integrate), Weaviate (hybrid search support), Milvus (scales to very large catalogues) and pgvector (a PostgreSQL extension, convenient when Postgres is already in the stack). The choice depends on catalogue size, latency requirements and infrastructure.

What is ANN and why not search exactly?

ANN is approximate nearest neighbour search. Exact search across every embedding in a catalogue costs O(n) per query — unrealistic at a million products with a query every 100 milliseconds. ANN indexes such as HNSW and IVF give up a fraction of a percent of accuracy in exchange for O(log n) lookup: tens of milliseconds instead of seconds.