Related: AI Agents for Business, Explained
Modern AI apps need a fast way to find things by meaning, not just by exact words. A vector database is the tool that makes this possible. It stores data as lists of numbers called vectors, and it finds the closest matches in milliseconds. This guide explains what a vector database is in plain language. It covers embeddings, similarity search, and the HNSW index, and shows when you actually need one.
Key takeaways
- A vector database stores embeddings, which are numeric fingerprints of text, images, or audio.
- It finds items by meaning using similarity search, not by matching exact keywords.
- The HNSW index makes search fast on millions of vectors by walking a smart graph.
- You need one mainly for semantic search, recommendations, and retrieval for AI chatbots.
- Popular options include Pinecone, Weaviate, Qdrant, Milvus, and the pgvector extension for Postgres.
- Small projects can start with pgvector. Large scale or heavy traffic favors a dedicated engine.
What is an embedding
An embedding is a list of numbers that describes the meaning of some content. A model reads your text and turns it into this list. The list often has hundreds or thousands of numbers. Each number is one dimension. Together they place the content as a point in a large space.
The key idea is simple. Content with similar meaning lands close together in this space. The words car and automobile sit near each other. The word banana sits far away. This is true even when the exact words differ. That is why embeddings power search by meaning.
You create embeddings with a model. The same model must be used for both your stored data and your search queries. If the models differ, the numbers will not line up and results will be wrong.
What is similarity search
Once your data is stored as vectors, search becomes a math problem. You turn the user query into a vector with the same model. Then you ask the database for the stored vectors that are nearest to it. Nearest means most similar in meaning.
| Metric | What it measures | Common use |
|---|---|---|
| Cosine similarity | The angle between two vectors | Text search, most popular choice |
| Dot product | Alignment and length together | Recommendations, ranking |
| Euclidean distance | Straight line gap between points | Image and spatial data |
The database returns the top matches, for example the top ten. This is called k nearest neighbors search. On a few thousand items you could check every vector one by one. But on millions of items, checking everything is too slow. This is where a special index helps.
How HNSW indexes work
HNSW stands for Hierarchical Navigable Small World. It is the most common index in vector databases. You do not need the math to use it, but the idea is easy to picture.
Think of a map with several layers stacked on top of each other. The top layer has very few points and long links between them. The bottom layer has every point with short links to close neighbors. Search starts at the top layer and takes big jumps to get near the target fast. Then it drops down a layer and takes smaller steps until it reaches the bottom and finds the closest points.
This layered approach is very fast. It does not check every vector. The trade off is that HNSW gives approximate results, not always the perfect match. In practice the accuracy is high enough for almost all apps. You can tune it with two main settings:
- A build setting that controls how many links each point keeps. More links give better accuracy but use more memory.
- A search setting that controls how wide the search explores. A wider search is more accurate but a little slower.
When you need a vector database
You do not need a vector database for every app. A normal database with keyword search is fine for many tasks. Reach for a vector database when meaning matters more than exact words. Here are clear cases:
- Semantic search across documents, support tickets, or product catalogs.
- Retrieval for an AI chatbot, so it answers from your own content. This is the core of retrieval augmented generation, often called RAG.
- Recommendations, such as items like this one or similar articles.
- Duplicate detection and clustering of large text or image sets.
- Image or audio search by content rather than by file name.
To see how this fits a full AI stack, read our guide on production RAG architecture. If you are deciding how to teach a model your data, our note on RAG versus fine tuning is a good next step.
Popular options to consider
Several engines are mature and widely used. The right pick depends on scale, budget, and whether you want a managed service or self hosting.
| Option | Type | Good fit for |
|---|---|---|
| pgvector | Postgres extension | Teams already on Postgres who want one database |
| Pinecone | Managed cloud service | Fast start with no servers to run |
| Weaviate | Open source, managed option | Built in modules and hybrid search |
| Qdrant | Open source, managed option | Strong filtering and easy self hosting |
| Milvus | Open source, built for scale | Very large datasets and high throughput |
A common path is to start with pgvector inside an existing Postgres database. When traffic, data size, or feature needs grow, you move to a dedicated engine. Many of these tools also support hybrid search, which mixes keyword and vector search for better results.
FAQ
Is a vector database a replacement for my normal database?
No. It works alongside your main database. Your normal database still holds records, users, and orders. The vector database holds embeddings and handles search by meaning. Many teams keep both and sync the data between them.
Do I need machine learning skills to use one?
Not much. You call an embedding model to turn content into vectors, then store and query them. The model and the database handle the hard parts. Most of the work is normal backend code, plus care to use the same model on both sides.
How many vectors can these handle?
From a few thousand to billions, depending on the engine and hardware. Small sets run fine on a single server. Very large sets need engines built for scale, such as Milvus, and may use sharding across machines. Memory is usually the main limit with HNSW.
Working with Apex Logic
We build AI apps, chatbots, and search features that use vector databases the right way. We help you pick a sensible option, design the embedding pipeline, and keep search fast and accurate as you grow. If you want semantic search or an AI assistant trained on your own content, see our AI solutions or contact us.
References
Malkov and Yashunin, research paper introducing the HNSW algorithm.
pgvector project documentation.
Vendor documentation from Pinecone, Weaviate, Qdrant, and Milvus.
Comments