GeraudBourdin / doctrine-pgvector

Postgresql vector type with doctrine

PGVector for Doctrine

Description

PGVector type for Doctrine

Installation

composer require partitech/doctrine-pgvector

Configuration Doctrine

doctrine:
  dbal:
    types:
      vector: Partitech\DoctrinePgVector\Type\VectorType
  orm:
    dql:
      string_functions:
        distance: Partitech\DoctrinePgVector\Query\Distance
        inner_product: Partitech\DoctrinePgVector\Query\InnerProduct
        cosine_similarity: Partitech\DoctrinePgVector\Query\CosineSimilarity

Utilisation

You can now use vector type in your entities :

use Doctrine\ORM\Mapping as ORM;

/**
* @ORM\Entity()
  */
  class YourEntity
  {

    #[ORM\Column(type: 'vector', length: 1024, nullable: true)]
    private $vectors;
    
  }

If you use symfony console make:entity add manually the length parameter attribute as vector. Length is your model embedding's dimension.

For example OpenAi use these dimensions:

text-embedding-3-small : 1536

text-embedding-3-large : 3072 (customizable)

Mistral AI

Mistral-embed : 1024

Additionally, you should manually add an HNSW index to your vector's column. Be aware that dimension should be 2000 max for HNSW indexes.

L2 distance

CREATE INDEX ON items USING hnsw (embedding vector_l2_ops);

Inner product

CREATE INDEX ON items USING hnsw (embedding vector_ip_ops);

Cosine distance

CREATE INDEX ON items USING hnsw (embedding vector_cosine_ops);

Basic usage:

distance

To get

SELECT * FROM embeddings WHERE vectors <-> '[3,1,2]' < 5

use

$floatArray = array_map(function() {
    return mt_rand(0, 1000000) / 1000000;
}, array_fill(0, 1024, null));

$query = $this->entityManager->createQuery(
    "SELECT i FROM App\Entity\Embeddings i ORDER BY distance(i.vectors, :vector) ASC"
);
$query->setParameter('vector', $floatArray, 'vector');
$results = $query->setMaxResults(5)->getResult();
dump($results);

$qb = $this->entityManager->createQueryBuilder();
$qb->select('e')
    ->from('App:Embeddings', 'e')
    ->orderBy('distance(e.vectors, :vector)')
    ->setParameter('vector', $floatArray, 'vector')
    ->setMaxResults(5)
    ;
$result = $qb->getQuery()->getResult();
dump($result);

Inner product

To get

SELECT (vectors <#> '[3,1,2]') * -1, * FROM embeddings

use

$floatArray = array_map(function() {
    return mt_rand(0, 1000000) / 1000000;
}, array_fill(0, 1024, null));

$query = $this->entityManager->createQuery(
    "SELECT inner_product(e.vectors, :vector) , e FROM App\Entity\Embeddings e"
);
$query->setParameter('vector', $floatArray, 'vector');
$results = $query->setMaxResults(5)->getResult();
dump($results);

$qb = $this->entityManager->createQueryBuilder();
$qb->select('e')
    ->addSelect('inner_product(e.vectors, :vector)')
    ->from('App:Embeddings', 'e')
    ->setParameter('vector', $floatArray, 'vector')
    ->setMaxResults(5)
    ;
$result = $qb->getQuery()->getResult();
dump($result);

Cosine similarity

To get

SELECT 1 - (vectors <=> '[3,1,2]'), * FROM embeddings

use

$floatArray = array_map(function() {
return mt_rand(0, 1000000) / 1000000;
}, array_fill(0, 1024, null));

$query = $this->entityManager->createQuery(
    "SELECT cosine_similarity(e.vectors, :vector) , e FROM App\Entity\Embeddings e"
);
$query->setParameter('vector', $floatArray, 'vector');
$results = $query->setMaxResults(5)->getResult();
dump($results);

$qb = $this->entityManager->createQueryBuilder();
$qb->select('e')
    ->addSelect('cosine_similarity(e.vectors, :vector)')
    ->from('App:Embeddings', 'e')
    ->setParameter('vector', $floatArray, 'vector')
    ->setMaxResults(5)
    ;
$result = $qb->getQuery()->getResult();
dump($result);