What is a vector database and how does it work?

What this video shows

The question comes first. Nora runs a small home-renovation company. A customer calls: her shower is leaking, is it still under warranty? The answer is somewhere in two thousand files. Nora searches for “leak” and finds nothing, because the warranty paper never says leak. It says “water damage, from a failed seal”. Normal search looks for the same letters, so it misses different words with the same meaning.

Sam, a software engineer, explains the fix: a vector database searches by meaning, not by exact words. It works in five steps. One: every file is cut into small pieces, about a page each. Two: an AI model reads each piece and turns its meaning into a long list of numbers, often more than a thousand. That list is called a vector; think of it as an address on a map of meaning, where similar meanings sit close together. Three: the database keeps all these addresses in a smart index, like a library catalogue, so it stays fast. Four: the same AI model turns the question into an address on the same map. Five: the database takes the pieces closest to the question. The warranty paragraph is the closest, even though it never says leak.

Then the AI app gives just those few pieces to the AI model, which answers and shows the page it used. Sam is honest about the limit: for exact codes and names, such as a part number, keyword search is better, so good systems use both (hybrid search). And often no big new system is needed: normal databases such as Postgres can store vectors too. This is how I make the AI apps I build for daily work answer from a company’s own files.

What you will learn

• Why normal search misses an answer that uses different words with the same meaning.
• What a vector is: a long list of numbers that works like an address on a map of meaning.
• The five steps: cut the files into pieces, turn each piece into an address, store them with an index, give the question an address, take the nearest pieces.
• Why AI apps use a vector database: the AI model does not know your files and cannot read them all at once, so it gets the right few pages.
• The honest limits: exact codes and names need keyword search too (hybrid search), and your own database can often do the job.

The numbers behind the picture

The video gives only a few numbers, in words. The exact numbers and their sources are here.

• 2025: Big storage and database products added vector search as a normal feature. Amazon S3 Vectors became generally available on 2 December 2025: up to two billion vectors in one index, about 100 milliseconds for frequent searches, and up to 90% lower cost to upload, store and search. SQL Server 2025 shipped a vector data type the same month. (AWS, 2 December 2025; Microsoft, 18 November 2025)
• 2026: The common embedding models turn each piece of text into 1,024 to 3,072 numbers by default: OpenAI text-embedding-3-small 1,536 and text-embedding-3-large 3,072, Google Gemini 3,072, Cohere Embed v5 2,048, Voyage 4 1,024. PostgreSQL, which can store vectors with the free pgvector extension, is used by 57.9% of developers in the 2026 Stack Overflow survey. Analysts put the vector database market at about 3 to 3.7 billion US dollars this year; their numbers differ because each firm counts the market differently. (Vendor documentation, read 10 October 2026; Stack Overflow Developer Survey 2026; Grand View Research and Meticulous Research, 2026)
• 2027, my guess: Vector search becomes a normal part of the database a company already has, and mixing search by meaning with exact-word search becomes the default. Gartner predicts that by 2028, 80% of generative AI business apps will be built on companies’ existing data platforms. (my guess; Gartner, June 2025)

Questions people ask

What is the difference between a vector database and a normal database?

A normal database finds rows by exact values and words. A vector database stores the meaning of each piece of text as a list of numbers and finds the pieces closest in meaning to a question. In MongoDB’s words: “Unlike traditional full-text search which finds text matches, vector search finds vectors that are close to your search query in multi-dimensional space.” Today many normal databases can do both, for example PostgreSQL with pgvector, SQL Server 2025 and MongoDB Atlas.

What is an embedding?

It is the long list of numbers that an AI model makes from a piece of text, the “address on the map of meaning” in the video. OpenAI’s definition: “An embedding is a vector (list) of floating point numbers.” Texts with a similar meaning get similar numbers. One point to know: if you change the embedding model later, you must make the numbers again for all your files.

Is a vector database the same as RAG?

No. RAG is the whole method: find the right pages in your own files first, then let the AI answer from them. A vector database is the most common tool for the finding step. The video “What is RAG?” shows the whole method in one minute.

Why does it sometimes miss an exact part number or code?

Search by meaning is good at ideas, not at exact strings. Microsoft’s search documentation says product codes, special jargon, dates and people’s names “perform better with keyword search because it can identify exact matches.” That is why good systems run both searches and mix the results: hybrid search. In Anthropic’s own test, adding keyword search to search by meaning cut failed searches from 3.7% to 2.9%.

Do I need a separate vector database?

Often not. If you already use PostgreSQL, the free pgvector extension adds vector search to it, and many other databases now have it built in. For a very small set of files you may not need search at all: Anthropic notes that below about 500 pages you can give the whole text to the model.

Does the AI model behind it need to be the biggest one?

Not usually. Once the right few pages are found, answering from them is often a simple job. Start with a small model, test it on your real questions, and step up only if you need to. The video “Which AI model should I use?” shows this rule.

Further reading

• MongoDB, “Vector Search overview” (search by meaning, in plain words). mongodb.com
• Microsoft, “Integrated vector database” (what a vector database is, and using the one you already have). learn.microsoft.com
• OpenAI, “Embeddings” guide. developers.openai.com
• Google, Gemini “Embeddings”. ai.google.dev
• pgvector, vector search for PostgreSQL. github.com
• Microsoft, “Hybrid search overview” (meaning plus exact words). learn.microsoft.com
• Anthropic, “Introducing Contextual Retrieval”, September 2024. anthropic.com
• AWS, “Amazon S3 Vectors is now generally available”, December 2025. aws.amazon.com