README

Framework-agnostic Vertex AI client for PHP: on-demand Gemini/Imagen/Veo, Google Cloud TTS, Deepgram STT, and Vertex batch prediction jobs (JSONL + GCS).

Install

composer require kozmonos/vertex-ai

Published on Packagist. Path repositories are only needed when developing the package locally alongside a consuming app.

Configuration

Copy config/vertex-ai.php into your app or build VertexConfig::fromArray() manually:

use Kozmonos\VertexAi\Vertex\VertexConfig;

$config = VertexConfig::fromArray(require 'vertex-ai.php', basePath: '/app');

Required for on-demand Vertex calls:

vertex.project_id
vertex.credentials_path (service account JSON)

Required for batch:

vertex.batch.gcs_bucket

Outbound reference images

HTTP/HTTPS reference image downloads are denied unless you configure an allowlist:

'outbound' => [
    'allowed_hosts' => ['cdn.example.com'],
],

See SECURITY.md for the full hardening checklist.

On-demand text

use Kozmonos\VertexAi\Infrastructure\Auth\GoogleAuthTokenProvider;
use Kozmonos\VertexAi\Infrastructure\Http\VertexHttpTransport;
use Kozmonos\VertexAi\Support\AiModelRegistry;
use Kozmonos\VertexAi\Vertex\VertexTextProvider;

$registry = AiModelRegistry::fromDefaults();
$transport = new VertexHttpTransport($config, new GoogleAuthTokenProvider($config));
$text = new VertexTextProvider($transport, $registry);

$result = $text->generateText(
    systemPrompt: 'You are helpful.',
    userPrompt: 'Say hello.',
    modelId: 'vertex-gemini-3.1-pro',
);

On-demand video (Veo)

use Kozmonos\VertexAi\Support\HttpReferenceImageLoader;
use Kozmonos\VertexAi\Vertex\VertexVideoProvider;

$video = new VertexVideoProvider($transport, $registry, new HttpReferenceImageLoader($allowedHosts));

// Text → video
$result = $video->generateVideo(
    prompt: 'A red balloon drifting over a city',
    modelId: 'veo-3.1-fast-vertex',
    durationSeconds: 4,
    options: [
        'aspect_ratio' => '16:9',
        'storage_uri' => 'gs://my-bucket/veo-output/',
    ],
);

// Image → video: HTTP/MinIO/data URIs are loaded via ReferenceImageLoader
// (bytesBase64Encoded + mimeType). gs:// URIs are sent as gcsUri + mimeType.
$result = $video->generateVideo(
    prompt: 'Slow camera push-in',
    modelId: 'veo-3.1-fast-vertex',
    durationSeconds: 4,
    options: [
        'source_image_url' => 'https://cdn.example.com/frame.png',
        // or: 'source_image_url' => 'gs://bucket/refs/frame.png',
        // or: 'source_image_base64' => $base64, 'source_image_mime_type' => 'image/png',
    ],
);

Batch submit

use Kozmonos\VertexAi\Batch\Support\VertexBatchJsonl;
use Kozmonos\VertexAi\Infrastructure\Batch\FakeBatchJobGateway;
use Kozmonos\VertexAi\Infrastructure\Gcs\InMemoryGcsObjectStore;
use Kozmonos\VertexAi\Vertex\VertexBatchService;

$batch = new VertexBatchService($config, new InMemoryGcsObjectStore('my-bucket'), new FakeBatchJobGateway());

$submission = $batch->submitBatchJob(
    jsonlContent: VertexBatchJsonl::encode([['key' => 'line-1', 'request' => ['contents' => []]]]),
    modelResource: 'publishers/google/models/gemini-2.5-flash',
    displayName: 'my-batch',
);

Gemini image batch requests keep Vertex safety filters enabled by default. Pass disableSafetyFilters: true to VertexBatchRequestBuilder::geminiImage() only when you explicitly need them disabled.

Usage recording

Wrap providers with Kozmonos\VertexAi\Usage\Recording\* and implement Kozmonos\VertexAi\Contracts\UsageRecorder, or use NullUsageRecorder in tests.

Tests

composer install
composer test
composer analyse

License

MIT. See LICENSE.

kozmonos / vertex-ai

Maintainers

Package info

Statistics

Security