pgvector 0.4.2

pgvector support for Python

pgvector-python

pgvector support for Python

Supports Django, SQLAlchemy, SQLModel, Psycopg 3, Psycopg 2, asyncpg, pg8000, and Peewee

Installation

Run:

pip install pgvector

And follow the instructions for your database library:

• Django
• SQLAlchemy
• SQLModel
• Psycopg 3
• Psycopg 2
• asyncpg
• pg8000
• Peewee

Or check out some examples:

• Retrieval-augmented generation with Ollama
• Embeddings with OpenAI
• Binary embeddings with Cohere
• Sentence embeddings with SentenceTransformers
• Hybrid search with SentenceTransformers (Reciprocal Rank Fusion)
• Hybrid search with SentenceTransformers (cross-encoder)
• Sparse search with Transformers
• Late interaction search with ColBERT
• Visual document retrieval with ColPali
• Image search with PyTorch
• Image search with perceptual hashing
• Morgan fingerprints with RDKit
• Topic modeling with Gensim
• Implicit feedback recommendations with Implicit
• Explicit feedback recommendations with Surprise
• Recommendations with LightFM
• Horizontal scaling with Citus
• Bulk loading with COPY

Django

Create a migration to enable the extension

from pgvector.django import VectorExtension

class Migration(migrations.Migration):
    operations = [
        VectorExtension()
    ]

Add a vector field to your model

from pgvector.django import VectorField

class Item(models.Model):
    embedding = VectorField(dimensions=3)

Also supports HalfVectorField, BitField, and SparseVectorField

Insert a vector

item = Item(embedding=[1, 2, 3])
item.save()

Get the nearest neighbors to a vector

from pgvector.django import L2Distance

Item.objects.order_by(L2Distance('embedding', [3, 1, 2]))[:5]

Also supports MaxInnerProduct, CosineDistance, L1Distance, HammingDistance, and JaccardDistance

Get the distance

Item.objects.annotate(distance=L2Distance('embedding', [3, 1, 2]))

Get items within a certain distance

Item.objects.alias(distance=L2Distance('embedding', [3, 1, 2])).filter(distance__lt=5)

Average vectors

from django.db.models import Avg

Item.objects.aggregate(Avg('embedding'))

Also supports Sum

Add an approximate index

from pgvector.django import HnswIndex, IvfflatIndex

class Item(models.Model):
    class Meta:
        indexes = [
            HnswIndex(
                name='my_index',
                fields=['embedding'],
                m=16,
                ef_construction=64,
                opclasses=['vector_l2_ops']
            ),
            # or
            IvfflatIndex(
                name='my_index',
                fields=['embedding'],
                lists=100,
                opclasses=['vector_l2_ops']
            )
        ]

Use vector_ip_ops for inner product and vector_cosine_ops for cosine distance

Half-Precision Indexing

Index vectors at half-precision

from django.contrib.postgres.indexes import OpClass
from django.db.models.functions import Cast
from pgvector.django import HnswIndex, HalfVectorField

class Item(models.Model):
    class Meta:
        indexes = [
            HnswIndex(
                OpClass(Cast('embedding', HalfVectorField(dimensions=3)), name='halfvec_l2_ops'),
                name='my_index',
                m=16,
                ef_construction=64
            )
        ]

Note: Add 'django.contrib.postgres' to INSTALLED_APPS to use OpClass

Get the nearest neighbors

distance = L2Distance(Cast('embedding', HalfVectorField(dimensions=3)), [3, 1, 2])
Item.objects.order_by(distance)[:5]

SQLAlchemy

Enable the extension

session.execute(text('CREATE EXTENSION IF NOT EXISTS vector'))

Add a vector column

from pgvector.sqlalchemy import Vector

class Item(Base):
    embedding = mapped_column(Vector(3))

Also supports HALFVEC, BIT, and SPARSEVEC

Insert a vector

item = Item(embedding=[1, 2, 3])
session.add(item)
session.commit()

Get the nearest neighbors to a vector

session.scalars(select(Item).order_by(Item.embedding.l2_distance([3, 1, 2])).limit(5))

Also supports max_inner_product, cosine_distance, l1_distance, hamming_distance, and jaccard_distance

Get the distance

session.scalars(select(Item.embedding.l2_distance([3, 1, 2])))

Get items within a certain distance

session.scalars(select(Item).filter(Item.embedding.l2_distance([3, 1, 2]) < 5))

Average vectors

from pgvector.sqlalchemy import avg

session.scalars(select(avg(Item.embedding))).first()

Also supports sum

Add an approximate index

index = Index(
    'my_index',
    Item.embedding,
    postgresql_using='hnsw',
    postgresql_with={'m': 16, 'ef_construction': 64},
    postgresql_ops={'embedding': 'vector_l2_ops'}
)
# or
index = Index(
    'my_index',
    Item.embedding,
    postgresql_using='ivfflat',
    postgresql_with={'lists': 100},
    postgresql_ops={'embedding': 'vector_l2_ops'}
)

index.create(engine)

Use vector_ip_ops for inner product and vector_cosine_ops for cosine distance

Half-Precision Indexing

Index vectors at half-precision

from pgvector.sqlalchemy import HALFVEC
from sqlalchemy.sql import func

index = Index(
    'my_index',
    func.cast(Item.embedding, HALFVEC(3)).label('embedding'),
    postgresql_using='hnsw',
    postgresql_ops={'embedding': 'halfvec_l2_ops'}
)

Get the nearest neighbors

order = func.cast(Item.embedding, HALFVEC(3)).l2_distance([3, 1, 2])
session.scalars(select(Item).order_by(order).limit(5))

Binary Quantization

Use expression indexing for binary quantization

from pgvector.sqlalchemy import BIT
from sqlalchemy.sql import func

index = Index(
    'my_index',
    func.cast(func.binary_quantize(Item.embedding), BIT(3)).label('embedding'),
    postgresql_using='hnsw',
    postgresql_ops={'embedding': 'bit_hamming_ops'}
)

Get the nearest neighbors by Hamming distance

order = func.cast(func.binary_quantize(Item.embedding), BIT(3)).hamming_distance(func.binary_quantize(func.cast([3, -1, 2], VECTOR(3))))
session.scalars(select(Item).order_by(order).limit(5))

Re-rank by the original vectors for better recall

order = func.cast(func.binary_quantize(Item.embedding), BIT(3)).hamming_distance(func.binary_quantize(func.cast([3, -1, 2], VECTOR(3))))
subquery = session.query(Item).order_by(order).limit(20).subquery()
session.scalars(select(subquery).order_by(subquery.c.embedding.cosine_distance([3, -1, 2])).limit(5))

Arrays

Add an array column

from pgvector.sqlalchemy import Vector
from sqlalchemy import ARRAY

class Item(Base):
    embeddings = mapped_column(ARRAY(Vector(3)))

And register the types with the underlying driver

For Psycopg 3, use

from pgvector.psycopg import register_vector
from sqlalchemy import event

@event.listens_for(engine, "connect")
def connect(dbapi_connection, connection_record):
    register_vector(dbapi_connection)

For async connections with Psycopg 3, use

from pgvector.psycopg import register_vector_async
from sqlalchemy import event

@event.listens_for(engine.sync_engine, "connect")
def connect(dbapi_connection, connection_record):
    dbapi_connection.run_async(register_vector_async)

For Psycopg 2, use

from pgvector.psycopg2 import register_vector
from sqlalchemy import event

@event.listens_for(engine, "connect")
def connect(dbapi_connection, connection_record):
    register_vector(dbapi_connection, arrays=True)

SQLModel

Enable the extension

session.exec(text('CREATE EXTENSION IF NOT EXISTS vector'))

Add a vector column

from pgvector.sqlalchemy import Vector

class Item(SQLModel, table=True):
    embedding: Any = Field(sa_type=Vector(3))

Also supports HALFVEC, BIT, and SPARSEVEC

Insert a vector

item = Item(embedding=[1, 2, 3])
session.add(item)
session.commit()

Get the nearest neighbors to a vector

session.exec(select(Item).order_by(Item.embedding.l2_distance([3, 1, 2])).limit(5))

Also supports max_inner_product, cosine_distance, l1_distance, hamming_distance, and jaccard_distance

Get the distance

session.exec(select(Item.embedding.l2_distance([3, 1, 2])))

Get items within a certain distance

session.exec(select(Item).filter(Item.embedding.l2_distance([3, 1, 2]) < 5))

Average vectors

from pgvector.sqlalchemy import avg

session.exec(select(avg(Item.embedding))).first()

Also supports sum

Add an approximate index

from sqlmodel import Index

index = Index(
    'my_index',
    Item.embedding,
    postgresql_using='hnsw',
    postgresql_with={'m': 16, 'ef_construction': 64},
    postgresql_ops={'embedding': 'vector_l2_ops'}
)
# or
index = Index(
    'my_index',
    Item.embedding,
    postgresql_using='ivfflat',
    postgresql_with={'lists': 100},
    postgresql_ops={'embedding': 'vector_l2_ops'}
)

index.create(engine)

Use vector_ip_ops for inner product and vector_cosine_ops for cosine distance

Psycopg 3

Enable the extension

conn.execute('CREATE EXTENSION IF NOT EXISTS vector')

Register the types with your connection

from pgvector.psycopg import register_vector

register_vector(conn)

For connection pools, use

def configure(conn):
    register_vector(conn)

pool = ConnectionPool(..., configure=configure)

For async connections, use

from pgvector.psycopg import register_vector_async

await register_vector_async(conn)

Create a table

conn.execute('CREATE TABLE items (id bigserial PRIMARY KEY, embedding vector(3))')

Insert a vector

embedding = np.array([1, 2, 3])
conn.execute('INSERT INTO items (embedding) VALUES (%s)', (embedding,))

Get the nearest neighbors to a vector

conn.execute('SELECT * FROM items ORDER BY embedding <-> %s LIMIT 5', (embedding,)).fetchall()

Add an approximate index

conn.execute('CREATE INDEX ON items USING hnsw (embedding vector_l2_ops)')
# or
conn.execute('CREATE INDEX ON items USING ivfflat (embedding vector_l2_ops) WITH (lists = 100)')

Use vector_ip_ops for inner product and vector_cosine_ops for cosine distance

Psycopg 2

Enable the extension

cur = conn.cursor()
cur.execute('CREATE EXTENSION IF NOT EXISTS vector')

Register the types with your connection or cursor

from pgvector.psycopg2 import register_vector

register_vector(conn)

Create a table

cur.execute('CREATE TABLE items (id bigserial PRIMARY KEY, embedding vector(3))')

Insert a vector

embedding = np.array([1, 2, 3])
cur.execute('INSERT INTO items (embedding) VALUES (%s)', (embedding,))

Get the nearest neighbors to a vector

cur.execute('SELECT * FROM items ORDER BY embedding <-> %s LIMIT 5', (embedding,))
cur.fetchall()

Add an approximate index

cur.execute('CREATE INDEX ON items USING hnsw (embedding vector_l2_ops)')
# or
cur.execute('CREATE INDEX ON items USING ivfflat (embedding vector_l2_ops) WITH (lists = 100)')

Use vector_ip_ops for inner product and vector_cosine_ops for cosine distance

asyncpg

Enable the extension

await conn.execute('CREATE EXTENSION IF NOT EXISTS vector')

Register the types with your connection

from pgvector.asyncpg import register_vector

await register_vector(conn)

or your pool

async def init(conn):
    await register_vector(conn)

pool = await asyncpg.create_pool(..., init=init)

Create a table

await conn.execute('CREATE TABLE items (id bigserial PRIMARY KEY, embedding vector(3))')

Insert a vector

embedding = np.array([1, 2, 3])
await conn.execute('INSERT INTO items (embedding) VALUES ($1)', embedding)

Get the nearest neighbors to a vector

await conn.fetch('SELECT * FROM items ORDER BY embedding <-> $1 LIMIT 5', embedding)

Add an approximate index

await conn.execute('CREATE INDEX ON items USING hnsw (embedding vector_l2_ops)')
# or
await conn.execute('CREATE INDEX ON items USING ivfflat (embedding vector_l2_ops) WITH (lists = 100)')

Use vector_ip_ops for inner product and vector_cosine_ops for cosine distance

pg8000

Enable the extension

conn.run('CREATE EXTENSION IF NOT EXISTS vector')

Register the types with your connection

from pgvector.pg8000 import register_vector

register_vector(conn)

Create a table

conn.run('CREATE TABLE items (id bigserial PRIMARY KEY, embedding vector(3))')

Insert a vector

embedding = np.array([1, 2, 3])
conn.run('INSERT INTO items (embedding) VALUES (:embedding)', embedding=embedding)

Get the nearest neighbors to a vector

conn.run('SELECT * FROM items ORDER BY embedding <-> :embedding LIMIT 5', embedding=embedding)

Add an approximate index

conn.run('CREATE INDEX ON items USING hnsw (embedding vector_l2_ops)')
# or
conn.run('CREATE INDEX ON items USING ivfflat (embedding vector_l2_ops) WITH (lists = 100)')

Use vector_ip_ops for inner product and vector_cosine_ops for cosine distance

Peewee

Add a vector column

from pgvector.peewee import VectorField

class Item(BaseModel):
    embedding = VectorField(dimensions=3)

Also supports HalfVectorField, FixedBitField, and SparseVectorField

Insert a vector

item = Item.create(embedding=[1, 2, 3])

Get the nearest neighbors to a vector

Item.select().order_by(Item.embedding.l2_distance([3, 1, 2])).limit(5)

Also supports max_inner_product, cosine_distance, l1_distance, hamming_distance, and jaccard_distance

Get the distance

Item.select(Item.embedding.l2_distance([3, 1, 2]).alias('distance'))

Get items within a certain distance

Item.select().where(Item.embedding.l2_distance([3, 1, 2]) < 5)

Average vectors

from peewee import fn

Item.select(fn.avg(Item.embedding).coerce(True)).scalar()

Also supports sum

Add an approximate index

Item.add_index('embedding vector_l2_ops', using='hnsw')

Use vector_ip_ops for inner product and vector_cosine_ops for cosine distance

Reference

Half Vectors

Create a half vector from a list

vec = HalfVector([1, 2, 3])

Or a NumPy array

vec = HalfVector(np.array([1, 2, 3]))

Get a list

lst = vec.to_list()

Get a NumPy array

arr = vec.to_numpy()

Sparse Vectors

Create a sparse vector from a list

vec = SparseVector([1, 0, 2, 0, 3, 0])

Or a NumPy array

vec = SparseVector(np.array([1, 0, 2, 0, 3, 0]))

Or a SciPy sparse array

arr = coo_array(([1, 2, 3], ([0, 2, 4],)), shape=(6,))
vec = SparseVector(arr)

Or a dictionary of non-zero elements

vec = SparseVector({0: 1, 2: 2, 4: 3}, 6)

Note: Indices start at 0

Get the number of dimensions

dim = vec.dimensions()

Get the indices of non-zero elements

indices = vec.indices()

Get the values of non-zero elements

values = vec.values()

Get a list

lst = vec.to_list()

Get a NumPy array

arr = vec.to_numpy()

Get a SciPy sparse array

arr = vec.to_coo()

History

View the changelog

Contributing

Everyone is encouraged to help improve this project. Here are a few ways you can help:

• Report bugs
• Fix bugs and submit pull requests
• Write, clarify, or fix documentation
• Suggest or add new features

To get started with development:

git clone https://github.com/pgvector/pgvector-python.git
cd pgvector-python
pip install -r requirements.txt
createdb pgvector_python_test
pytest

To run an example:

cd examples/loading
pip install -r requirements.txt
createdb pgvector_example
python3 example.py