[Platform] Add Ollama embeddings

[Guikingone]

Q	A
Bug fix?	no
New feature?	yes
Docs?	yes
Issues	None
License	MIT

Hi 👋🏻

This PR aims to add the support for embeddings via Ollama, the platform is updated to handle the new embedding client.

[Guikingone]

Ok, should be better 😅

[JoshuaBehrens]

Looks like I need to be faster with proposing my changes :D glad that you already have tests at hand

This is my example to proof, that Ollama can solve the same movie selection task

<?php

/*
 * This file is part of the Symfony package.
 *
 * (c) Fabien Potencier <[email protected]>
 *
 * For the full copyright and license information, please view the LICENSE
 * file that was distributed with this source code.
 */

use Symfony\AI\Agent\Agent;
use Symfony\AI\Agent\Toolbox\AgentProcessor;
use Symfony\AI\Agent\Toolbox\Tool\SimilaritySearch;
use Symfony\AI\Agent\Toolbox\Toolbox;
use Symfony\AI\Fixtures\Movies;
use Symfony\AI\Platform\Bridge\Meta\Llama;
use Symfony\AI\Platform\Bridge\Ollama\Embeddings;
use Symfony\AI\Platform\Bridge\Ollama\PlatformFactory;
use Symfony\AI\Platform\Message\Message;
use Symfony\AI\Platform\Message\MessageBag;
use Symfony\AI\Store\Document\Metadata;
use Symfony\AI\Store\Document\TextDocument;
use Symfony\AI\Store\Document\Vectorizer;
use Symfony\AI\Store\Indexer;
use Symfony\AI\Store\InMemoryStore;
use Symfony\Component\Uid\Uuid;

require_once dirname(__DIR__).'/bootstrap.php';

// initialize the store
$store = new InMemoryStore();

// create embeddings and documents
foreach (Movies::all() as $i => $movie) {
    $documents[] = new TextDocument(
        id: Uuid::v4(),
        content: 'Title: '.$movie['title'].\PHP_EOL.'Director: '.$movie['director'].\PHP_EOL.'Description: '.$movie['description'],
        metadata: new Metadata($movie),
    );
}

// create embeddings for documents
$platform = PlatformFactory::create(env('OLLAMA_HOST_URL'), http_client());
$vectorizer = new Vectorizer($platform, $embeddings = new Embeddings(Embeddings::MXBAI_EMBED_LARGE));
$indexer = new Indexer($vectorizer, $store, logger());
$indexer->index($documents);

$model = new Llama(Llama::V3_2_1B);

$similaritySearch = new SimilaritySearch($platform, $embeddings, $store);
$toolbox = new Toolbox([$similaritySearch], logger: logger());
$processor = new AgentProcessor($toolbox);
$agent = new Agent($platform, $model, [$processor], [$processor], logger());

$messages = new MessageBag(
    Message::forSystem('Please answer all user questions only using SimilaritySearch function.'),
    Message::ofUser('Which movie fits the theme of the mafia?')
);
$result = $agent->call($messages);

echo $result->getContent().\PHP_EOL;

[Guikingone]

Hi @JoshuaBehrens 👋🏻

Ah, didn't know that you were working on it, sorry 😅

Regarding your example, I just listed the model that are commonly used, if you have any other model in mind, feel free to list them, we can add them in the class 🙂

BTW, I agree (if the suggestion is to add an example via Ollama), we should add one with the Ollama stack

[JoshuaBehrens]

No worries @Guikingone I did not have any real dips on that ^^'

Implementation looks similar:

class Embeddings extends Model
{
    public const MXBAI_EMBED_LARGE = 'mxbai-embed-large';
    public const NOMIC_EMBED_TEXT = 'nomic-embed-text';
    public const ALL_MINILM = 'all-minilm';

    /**
     * @param array<string, mixed> $options
     */
    public function __construct(string $name = self::ALL_MINILM, array $options = [])
    {
        parent::__construct($name, [], $options);
    }
}

---

Right now I am refactoring the use of the modal class so I wanted to wait for my refactoring but I will just amend yours later :)

[chr-hertel]

Hey @Guikingone, thanks for working on that and extending our Ollama support!

My initial thought was that this is straight forward and a matching implementation, but checking it out and testing it got me thinking.

I think it would be great, if we can ease the Ollama usage a bit and reduce it to one model class and client. For deciding on the endpoint and result converter, we could use the model's capabilities. Either hard code them in the model class or dynamically fetch them from /api/show endpoint.

how does that sound?

[chr-hertel]

I don't think this addition is meaningful, since we create the embeddings also with models and the main feature of ollama is running models locally

[Guikingone]

Agree on the refactoring, the show endpoint could be a solution, I'll take a look on it.

[Guikingone]

Should be better, nice catch around the support method @chr-hertel 👍🏻

[chr-hertel]

Looks good & works now - thanks! 👍