Embeddings
AI Fundamentals
Embeddings are lists of numbers that capture aspects of the meaning of text, images, or other data, so software can find items with similar meaning.
AI Agents
What Is a Vector Database?
A vector database is a database built to store embeddings (lists of numbers representing the meaning of text, images, or other content) and search them by similarity. Rather than only matching exact keywords, it finds items whose vectors are closest to a query’s vector, which usually means closest in meaning. This is called similarity or nearest-neighbor search. AI tools use vector databases to quickly retrieve relevant information, such as document passages or products similar to one being viewed.
Also called vector store, vector DB
Example
An online furniture retailer is building a “You might also like” feature. The team uses a multimodal embedding model (one that handles images and text) to turn each product’s photo and description into vectors. These go into a vector database with each item’s ID, category, price, and stock status, details known as metadata.
When a shopper views a walnut mid-century armchair, its vectors become the query. The database returns the most similar products, such as other mid-century chairs in warm wood tones, even when their listings use different words. Metadata filters, applied during the search, exclude out-of-stock items and keep prices close to the armchair’s. Matches can reflect style and product type, not just shared keywords, though their quality depends on the embedding model and on the product photos and descriptions.
How a vector database stores and searches
A vector database does two jobs: it stores vectors, and it finds the closest ones to a query.
- Storing: Before anything reaches the database, your application or data pipeline breaks content into pieces, such as a product listing, a paragraph from a policy, or an image. An embedding model turns each piece into a vector, called an embedding. (Some databases can call the embedding model for you.) The database saves each vector alongside the original content (or a pointer to it) and metadata such as source, date, category, and who is allowed to see it.
- Searching: A question or item is converted into a vector with the same embedding model. The database compares it to the stored vectors using a similarity measure (a formula that scores how close two vectors are) and returns the closest matches ranked by score.
Checking a query against every stored vector gets slow as a collection grows. So vector databases typically build an index for approximate nearest-neighbor search, which may occasionally miss a true closest match in exchange for much faster results. How much accuracy is traded for speed is usually adjustable. Metadata filters narrow results further, for example to one product line, one region, or only documents a given user has permission to view.
Its role in retrieval-augmented generation and semantic search
Vector databases are the retrieval layer behind many AI features. In a common retrieval-augmented generation (RAG) setup, a user’s question is turned into a vector, the database returns the most similar passages from your documents, and those passages are given to a language model so it can answer from your material instead of relying only on what it learned in training. Not every RAG system works this way. Some use keyword search or combine both.
In semantic search, the same lookup lets a search box match by meaning. A search for “how do I reset my login” can surface a help article titled “Recovering account access,” even though the wording is different.
A vector database does not decide what content you have or whether it is correct. That depends on the knowledge base or other sources it draws from. It also does not write answers. It finds close matches, and other parts of the system decide what to do with them. The most similar passage is not always the most useful one, and if the source content is outdated or messy, a vector database will simply retrieve outdated or messy content faster.
Dedicated vector databases vs. built-in vector search
There are two broad ways to get vector search:
- Dedicated vector databases are designed specifically for storing and searching vectors. Teams tend to consider them for very large collections, heavy query volume, or when they want specialized indexing and tuning options.
- Vector search features in existing systems add vector storage and similarity search to databases, search engines, and data platforms that many businesses already run. This can keep vectors next to the rest of your business data and avoid managing one more system.
Neither approach is better in every case. The right fit depends on how much content you have, how fast results need to be, which systems your team already maintains, and how you handle security and access control. For smaller projects, the vector features of a platform you already use are often enough.
Do you need to pick a vector database?
Often, no. Many AI assistants and business tools that let you upload files or connect a document library handle embeddings and retrieval behind the scenes. When you add files to an AI assistant’s project or connect a help center to a support chatbot, vector search may be doing the work without you ever seeing it.
You are more likely to evaluate a vector database directly when building a custom AI application, such as an internal assistant that searches thousands of documents or a recommendation feature on your website. In that case, ask practical questions:
- Can it filter results by metadata such as date, product line, or customer?
- Can it respect permissions so people only retrieve content they are allowed to see?
- How do updates and deletions in your source documents reach the database?
- What happens if you switch embedding models? Vectors from different models are not comparable, so content usually has to be re-embedded.
- Where is the data stored, and does that meet your privacy and compliance requirements?
Further reading
- AI ImplementationUseful if you are planning a custom AI application that needs to connect to your own systems and data.
- AI Tool LibraryBrowse AI tools sorted by the job they do if you would rather use an existing tool than build retrieval yourself.
From definition to deployment
See an AI agent doing real work.
SN3 AI Employees are role-trained AI systems deployed inside the tools your team already uses, with guardrails and support after launch.