Manage unstructured vector stores in Postgres.

Supabase provides a Python client called vecs for managing unstructured vector stores. This client provides a set of useful tools for creating and querying collections in Postgres using the pgvector extension.

To see how Vecs works, use a local database. Make sure you have the Supabase CLI installed on your machine.

Initialize your project#

Start a local Postgres instance in any folder using the init and start commands. Make sure you have Docker running!

1
# Initialize your project
2
supabase init
3
4
# Start Postgres
5
supabase start

Create a collection#

Inside a Python shell, run the following commands to create a new collection called "docs", with 3 dimensions.

1
import vecs
2
3
# create vector store client
4
vx = vecs.create_client("postgresql://postgres:postgres@localhost:54322/postgres")
5
6
# create a collection of vectors with 3 dimensions
7
docs = vx.get_or_create_collection(name="docs", dimension=3)

Add embeddings#

Now we can insert some embeddings into our "docs" collection using the upsert() command:

1
import vecs
2
3
# create vector store client
4
docs = vecs.get_or_create_collection(name="docs", dimension=3)
5
6
# a collection of vectors with 3 dimensions
7
vectors=[
8
  ("vec0", [0.1, 0.2, 0.3], {"year": 1973}),
9
  ("vec1", [0.7, 0.8, 0.9], {"year": 2012})
10
]
11
12
# insert our vectors
13
docs.upsert(vectors=vectors)

Query the collection#

You can now query the collection to retrieve a relevant match:

1
import vecs
2
3
docs = vecs.get_or_create_collection(name="docs", dimension=3)
4
5
# query the collection filtering metadata for "year" = 2012
6
docs.query(
7
    data=[0.4,0.5,0.6],      # required
8
    limit=1,                         # number of records to return
9
    filters={"year": {"$eq": 2012}}, # metadata filters
10
)

For a more in-depth guide on vecs collections, see API.

Official Vecs Documentation: https://supabase.github.io/vecs/api
Source Code: https://github.com/supabase/vecs