🧲Vector Database

NEW

Also known as: Vector Store

A database optimized for storing, indexing, and querying high-dimensional embeddings using similarity search techniques.

DESCRIPTION

Vector databases manage dense numerical embeddings and support efficient similarity search at scale. They power semantic search, recommendation, and retrieval-augmented generation by indexing vectors with structures like HNSW or IVF. Many support approximate nearest neighbor for speed, plus metadata filtering and hybrid lexical–vector search. They integrate with embedding models and LLM pipelines. Design considerations include index type, recall–latency trade-offs, dimensionality, filtering, and durability. They complement, not replace, traditional transactional databases.

ALTERNATIVE DEFINITIONS

Milvus

A vector database is a database designed to store, index, and manage vector embeddings for efficient similarity search.

Stresses storage and indexing for similarity queries.

Source

Pinecone

A vector database is purpose-built to store and query vector embeddings and to perform fast nearest neighbor search.

Emphasizes nearest neighbor search speed.

Source

Weaviate

A vector database stores data objects and their vector representations to enable semantic search and machine learning applications.

Highlights semantic search and object–vector pairing.

Source

RELATED TERMS

Approximate Nearest Neighbor Vector Index Vector Embeddings Semantic Search HNSW FAISS Retrieval-Augmented Generation Metadata Filtering Hybrid Search ANN Benchmark

COMPETING TERMS

Relational Database Document Database

RELATIONS

• Broader: Database

KNOWLEDGE AREAS

Data Storage & Operations AI Data Engineering Data Services Platform & Tooling

TARGET AUDIENCE

AI Engineer Data Engineer MLOps Engineer Data Architect Cloud Architect

EXTERNAL SOURCES

• Wikipedia - Nearest neighbor search
• FAISS - Similarity search library widely used
• OpenSearch - k-NN and vector search plugin
• Milvus - Open-source vector database docs

TOP QUESTIONS

Why are Vector Databases relevant for LLM applications?

They store embeddings for semantic retrieval, enabling retrieval-augmented generation. This improves factual grounding and reduces hallucinations.

Category: Relevancy

What indexing options exist in Vector Databases?

Common indexes include HNSW, IVF, PQ/OPQ, and graph-based structures. They balance recall, speed, and memory.

Category: Overview

What is Approximate Nearest Neighbor (ANN) in Vector Databases?

ANN sacrifices exactness for latency, returning near-neighbors fast with high recall. It’s essential for large-scale search.

Category: Concept

How do Vector Databases differ from Search Engines?

Search engines focus on lexical matching; vector databases provide semantic similarity. Hybrid systems combine both for best results.

Category: Comparison

What are common mistakes with Vector Databases?

Skipping normalization and metric selection, ignoring metadata filters, and mis-sized indexes degrade recall and latency.

Category: Mistakes

What risks come with Vector Databases?

Drift in embeddings or model upgrades can break index compatibility. Privacy risks arise if embeddings leak sensitive content.

Category: Risks

How should I choose a distance metric?

Match metric to embedding training objective: cosine for normalized vectors, L2 for Euclidean-trained spaces, dot product for unnormalized.

Category: Best Practice

Can I use a Relational Database for embeddings?

Some support vector types and indexes, but specialized vector databases often provide better performance and tooling for large-scale workloads.

Category: Alternatives

How do I handle embedding model upgrades?

Version embeddings, dual-write during migration, and rebuild indexes gradually. Validate retrieval quality before cutover.

Category: Management