A vector represents data numerically in a way that captures its key characteristics, mapping it into a virtual mathematical space. In this space, similar items—like words, images, or objects—are positioned close together.

For example, consider the words “coat” and “jacket.” Traditional keyword-based searches would not connect these two words as similar, because their letters are quite different. An e-commerce system that wants to unite these keywords would need to do so manually. However, the vector representations of these two would be very close because they share similar meanings—delivering more accurate search results for users and saving time for developers.

Similarly, if you take two different pictures of cats, then pixel by pixel they might be vastly different. However, their vector embeddings would place them very close together in the mathematical space, just as a human would easily identify both of these as images of cats:

To make this work, an embedding model transforms raw data—such as images or text—into vector embeddings. pgvector stores these embeddings in your database. When a user submits a query, that input is also converted into a vector. pgvector then calculates the distance between the query vector and stored vectors to efficiently identify the "nearest neighbors" with the highest similarity scores.

Curious about different types of nearest neighbor searches? Check out our guide to generative AI app development.

To support today's AI-driven features, you need the ability to store and manage vector embeddings.

PostgreSQL can be powerful on its own, but because its data is rigidly structured into tables, rows, and columns, its query capability is largely limited to keyword and pattern matching.

In the world of AI, complex data like text, images, and audio is encoded as vector representations. These encodings enable AI models to grasp the context and semantic relationships within your data, forming the backbone of features like intelligent search, recommendations, and gen AI.

The pgvector extension brings semantic search to PostgreSQL, using vector embeddings to find results based on a query's meaning—rather than just keyword matches as SQL would. This process, known as similarity search, makes it straightforward to add advanced search capabilities directly into your applications without needing to re-architect or move data to a separate vector database.

Want to learn more about vector embeddings? Check out our guide to generative AI app development.

Cloud SQL and AlloyDB for PostgreSQL support pgvector, allowing you to store and query vector embeddings using standard SQL commands.

Use your preferred PostgreSQL client (such as psql, pgAdmin, or the Google Cloud console) to connect to your Cloud SQL or AlloyDB instance.

Run the following SQL command to enable the extension on your database. You only need to do this once per database.

Create a new table (or alter an existing one) to include a column for vector data. You must specify the dimensions of the vector. For example, to create a table for storing 3-dimensional embeddings:

You can insert vector embeddings just like standard data. Vectors are formatted as arrays enclosed in brackets.

You can now query your data to find the nearest neighbors. The <-> operator calculates Euclidean distance (L2 distance), which is commonly used to find the most similar items.

For larger datasets, adding an index can significantly speed up search performance. The HNSW and ScaNN indexes are commonly used options. Here’s an HNSW example: