Database

pgvector: Embeddings and vector similarity

pgvector is a Postgres extension for vector similarity search. It can also be used for storing embeddings.

Learn more about Supabase's AI & Vector offering.

Concepts

Vector similarity

Vector similarity refers to a measure of the similarity between two related items. For example, if you have a list of products, you can use vector similarity to find similar products. To do this, you need to convert each product into a "vector" of numbers, using a mathematical model. You can use a similar model for text, images, and other types of data. Once all of these vectors are stored in the database, you can use vector similarity to find similar items.

Embeddings

This is particularly useful if you're building on top of OpenAI's GPT-3. You can create and store embeddings for retrieval augmented generation.

Usage

Enable the extension

  1. Go to the Database page in the Dashboard.
  2. Click on Extensions in the sidebar.
  3. Search for "vector" and enable the extension.

Usage

Create a table to store vectors

1
2
3
4
5
6
create table posts (  id serial primary key,  title text not null,  body text not null,  embedding vector(384));

Storing a vector / embedding

In this example we'll generate a vector using Transformer.js, then store it in the database using the Supabase client.

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
import { pipeline } from '@xenova/transformers'const generateEmbedding = await pipeline('feature-extraction', 'Supabase/gte-small')const title = 'First post!'const body = 'Hello world!'// Generate a vector using Transformers.jsconst output = await generateEmbedding(body, {  pooling: 'mean',  normalize: true,})// Extract the embedding outputconst embedding = Array.from(output.data)// Store the vector in Postgresconst { data, error } = await supabase.from('posts').insert({  title,  body,  embedding,})

Specific usage cases

Queries with filtering

If you use an IVFFlat or HNSW index and naively filter the results based on the value of another column, you may get fewer rows returned than requested.

For example, the following query may return fewer than 5 rows, even if 5 corresponding rows exist in the database. This is because the embedding index may not return 5 rows matching the filter.

1
SELECT * FROM items WHERE category_id = 123 ORDER BY embedding <-> '[3,1,2]' LIMIT 5;

To get the exact number of requested rows, use iterative search to continue scanning the index until enough results are found.

More pgvector and Supabase resources