Vector Database

What is a vector database?

A vector database is a data system designed to store vector representations and retrieve the most similar items efficiently. It normally keeps each vector connected to its original content, identifier, and metadata.

Vector databases are commonly used for semantic search, RAG, recommendations, and AI-agent memory retrieval.

How a vector database works

During indexing, an embedding model converts documents or other items into vectors. The database stores those vectors and builds an index optimized for similarity search.

At query time:

1. The system creates an embedding for the query.
2. The database finds nearby stored vectors.
3. Metadata filters may restrict candidates.
4. The system returns matching content and relevance scores.
5. A reranker or application may refine the results.

For example, an agent can search product manuals by meaning and retrieve passages related to a user’s problem even when the wording differs.

Vector database versus ordinary database

Traditional databases are strong at exact values, structured filters, joins, and transactions. Vector databases specialize in nearest-neighbor search across high-dimensional vectors.

The approaches can work together. A retrieval request might use vector similarity for semantic relevance and structured fields for product, region, date, or permission checks. Some general databases also provide vector-search features, so “vector database” can describe a capability as well as a separate product category.

Vector database versus embedding model

The embedding model creates vectors. The database stores and searches them. Changing the model can make existing vectors incompatible or reduce quality, which may require rebuilding the index.

The database also does not verify whether a similar result is correct, current, or authorized.

Design and reliability concerns

Builders must choose an index method, similarity metric, metadata strategy, update process, and accuracy-latency trade-off. Approximate search can improve speed but may miss the true nearest results.

Stale vectors can remain after source documents change. Deleting the original document without removing its vector may expose outdated or unauthorized content. Access filters should be applied during retrieval, not only after results return.

Why it matters

A vector database makes semantic retrieval practical across large collections. It lets agents find relevant documents or memories without placing the entire collection into the context window.

Its value depends on the surrounding pipeline: good chunking, compatible embeddings, metadata, evaluation, reranking, and source governance. See [Memory in AI Agents](/memory-in-ai-agents/) for how vector stores support external agent memory.