Vector Database

A vector database stores numeric embeddings so that AI finds items by semantic meaning, not by keyword match. It powers semantic search, recommenders, and RAG pipelines — the core of modern generative AI, used by Netflix, Spotify, and enterprise chatbots and copilots.

Traditional relational databases were built for exact matches on structured fields like names, dates, and IDs. AI models produce a different beast: dense vectors of 384 to 3,072 dimensions per item, each capturing meaning learned from billions of training examples.

Comparing two vectors tells you how alike two ideas are, not whether two strings match. That single shift is why entire databases had to be rebuilt from scratch.

Pinecone, Weaviate, and Milvus emerged after 2019 as purpose-built stores for this data. They handle indexing, similarity search, metadata filters, and horizontal scaling in one system — the pieces AI teams used to stitch together by hand.

The trigger was generative AI. When ChatGPT went public in late 2022, teams building assistants and copilots needed to ground LLM responses in private data like documents, product catalogs, or chat history, so vector search became the only way to retrieve by meaning.

Data analytics teams tried to fit embeddings into Postgres, Elasticsearch, or Redis with add-ons and homegrown indexes. It worked at small scale, then broke at production traffic, which is why a dedicated tier appeared.

Pinecone’s technical overview frames the shift plainly: scalar databases fail at the complexity and scale of vector data. AWS documents the same conclusion, noting that k-NN indexes like HNSW and IVF are now core features for any modern search stack.

Key takeaways

• Stores high-dimensional embeddings that encode meaning, not literal keywords.
• Uses approximate nearest neighbor (ANN) algorithms like HNSW to search billions of vectors in milliseconds.
• Powers RAG, semantic search, recommendations, and multimodal AI systems.
• Emerged as dedicated infrastructure because relational databases struggle beyond a few million vectors.

How it works

A vector database works in three stages: an embedding model turns raw data into a numeric vector, an index groups similar vectors, and a query returns the closest matches using a distance metric such as cosine similarity.

Text, images, or audio pass through an embedding model such as OpenAI’s text-embedding-3, Cohere Embed, or Sentence-BERT. The output vector gets indexed with algorithms like HNSW, IVF, or product quantization.

At query time, the database compares the incoming query vector against indexed neighbors and returns the top-k closest matches, ranked by a distance function such as cosine similarity or dot product.

The choice of embedding model matters as much as the database. A 384-dimensional MiniLM vector uses one-quarter the storage of a 1,536-dimensional OpenAI vector, so at scale it changes cluster size, cost, and query latency. Most teams tune this trade-off empirically.

Approximate nearest neighbor (ANN) search is what makes this practical at scale. Exact k-NN over a billion 1,536-dimensional vectors would take hours.

HNSW cuts that to a few milliseconds by building a graph where similar vectors link directly, so the search walks a short path instead of scanning everything.

Metadata filters run alongside similarity search.

A query for “lightweight running shoes under $150” combines a vector match on “lightweight running shoes” with a scalar filter on price, so the database returns semantically relevant results that also meet the hard constraint.

Examples

Named companies deploy vector databases for search, personalization, and generative AI grounding. The pattern is the same: turn user or catalog data into vectors, store them, and query by similarity for real-time results.

Spotify rebuilt its podcast search on vector embeddings in 2022, using natural-language queries to surface episodes that keyword search missed. Retrieval latency stayed under 500 milliseconds across hundreds of millions of episodes.

Notion launched Notion AI in 2023 with a vector database backing its “Ask AI” feature, letting users query their own workspace by meaning rather than exact wording.

Shopify uses Pinecone to power semantic product search across millions of merchant catalogs, launched broadly in 2023 to lift discovery on long-tail queries.

Klarna’s shopping assistant, launched in 2024, uses a vector database to ground its LLM in real product data — cutting customer-service resolution time from 11 minutes to under 2, per its published benchmarks.

JPMorgan Chase disclosed in 2024 that its internal LLM Suite queries over 60,000 employees against a private vector database of research and policy documents, replacing hours of manual lookup with seconds.

BPO providers with data-labelling teams increasingly own the front of this pipeline: cleaning, tagging, and annotating the source data that feeds embedding models. High-quality inputs yield high-quality vectors, which yield useful retrieval at query time.

Poor labels quietly poison every downstream query, so the human review layer matters, often more than the model choice itself.

Related terms

Vector databases sit at the intersection of AI, search, and modern data engineering. The related terms below give the fuller picture of the stack around them and where each piece fits into a working pipeline.

• Artificial Intelligence: the broader field of building systems that reason from data.
• Machine Learning: the sub-discipline that trains embedding models on labeled examples.
• Generative AI: the class of models that ground their outputs on retrieved vectors.
• Natural Language Processing (NLP): the branch of AI focused on human language.
• Chatbot: a conversational interface that increasingly uses vectors for grounded answers.
• Data Science: the practice of turning raw data into features and models that vectors capture.
• Automation: the discipline of removing repetitive human steps that vectors help scale.

FAQ

What is a vector database used for?

It powers similarity search, recommendations, and retrieval-augmented generation. Any app that finds items by meaning, not exact match, including semantic search, image lookup, or grounding an LLM in private data, leans on a vector database for retrieval.

How is a vector database different from a traditional database?

Traditional databases match exact values in rows and columns. Vector databases match by similarity in high-dimensional space, using distance functions instead of equality checks. That shift lets them handle unstructured data like text, images, and audio.

What algorithms do vector databases use?

Most production systems rely on approximate nearest neighbor (ANN) methods like HNSW, IVF, and product quantization. These trade a small amount of accuracy for orders-of-magnitude speedup over exact search.

Do you need a vector database for RAG?

Not always. Proofs of concept run on in-memory FAISS or SQLite with pgvector, fine for hobby projects under 100,000 vectors. Production RAG, needing millions of documents, concurrent queries, and sub-100ms latency, needs dedicated vector databases.

Which vector databases lead the market in 2026?

Pinecone, Weaviate, Milvus, Qdrant, and Chroma dominate the dedicated space, while Postgres pgvector and Elasticsearch bring vector search into existing stacks.

See how outsourced data-labelling and human-in-the-loop review pair with vector databases at Outsource Accelerator.