laravel-ai-attributes

Add AI-powered computed attributes to any Eloquent model with a single trait.

class Article extends Model
{
    use HasAIAttributes;

    protected $aiAttributes = [
        'summary' => 'Summarize this in 2 sentences',
        'tags'    => 'Return 3-5 topic tags as JSON array',
    ];
}

$article = Article::find(1);

$article->ai_summary;  // → "Laravel 12 ships with..."
$article->ai_tags;     // → '["laravel", "php", "release-notes", ...]'

The first read calls the AI provider; subsequent reads with the same input come from cache.

Why?

You've probably written this code five times already:

• "Summarize this article"
• "Suggest tags for this post"
• "Translate this product description"
• "Generate a meta-description for SEO"

Every one of those is the same shape: take some model attributes, send them with a prompt, get text back, cache the result. This package collapses all of that into one trait.

Installation

composer require parselynk/laravel-ai-attributes

Publish the config:

php artisan vendor:publish --tag=ai-attributes-config

Set your API keys in .env:

ANTHROPIC_API_KEY=sk-ant-...
OPENAI_API_KEY=sk-...

# Or use Ollama — no key needed, runs locally:
OLLAMA_BASE_URL=http://localhost:11434
OLLAMA_MODEL=qwen2.5:7b

# Pick the default driver:
AI_ATTRIBUTES_DRIVER=claude   # or "openai" or "ollama"

Usage

1. Add the trait to a model

use Illuminate\Database\Eloquent\Model;
use Parselynk\AiAttributes\Concerns\HasAIAttributes;

class Article extends Model
{
    use HasAIAttributes;

    protected $aiAttributes = [
        'summary' => 'Summarize this article in 2 sentences.',
        'tags'    => 'Return 3 to 5 topic tags as a JSON array of strings.',
    ];
}

2. Read the AI attributes

Each key in $aiAttributes is exposed with an ai_ prefix:

$article = Article::find(1);

$article->ai_summary;   // calls the AI, cached on subsequent reads
$article->ai_tags;

3. Manually regenerate or invalidate

// Bypass the trait's magic and force a generation:
$article->generateAiAttribute('summary');

// Drop the cached value so the next read calls the AI again:
$article->forgetAiAttribute('summary');

How caching works

A SHA-256 cache key is built from:

• the model class (App\Models\Article)
• the attribute key (summary)
• the prompt (the string from $aiAttributes)
• the model attributes at read time (attributesToArray())

If any of those change, the value is regenerated. If none of them change, the AI is never called twice.

The cache uses your application's default cache store. Override per-app via .env:

AI_ATTRIBUTES_CACHE_ENABLED=true
AI_ATTRIBUTES_CACHE_STORE=redis
AI_ATTRIBUTES_CACHE_TTL=2592000   # 30 days, in seconds

Available drivers

Using the Ollama driver

Ollama lets you run open-source models locally. Great for privacy, cost control, or offline development.

1. Install and start Ollama:

brew install ollama          # macOS
brew services start ollama   # runs on http://localhost:11434

See ollama.com for other platforms.

2. Pull a model:

ollama pull qwen2.5:7b       # recommended — strong at structured output
# or
ollama pull llama3.2:3b      # smaller / faster, less consistent

3. Point the package at Ollama:

AI_ATTRIBUTES_DRIVER=ollama
OLLAMA_MODEL=qwen2.5:7b

That's it. Same trait, same caching, same retries — now talking to your local LLM.

Model recommendations for structured output (JSON, numbers, bool):

Tip: the package sends temperature: 0 to Ollama by default for predictable structured output. Override per-attribute or globally via OLLAMA_TEMPERATURE=0.7 if you want more creative text.

Pointing at a remote Ollama server (Docker, GPU box, etc.):

OLLAMA_BASE_URL=http://my-gpu-server:11434

Switch the default at runtime:

config(['ai-attributes.default' => 'openai']);

Adding a custom driver

The package uses Laravel's Manager pattern (the same one as Cache, Queue, Mail). Register a custom driver in any service provider:

use Parselynk\AiAttributes\AIManager;
use Parselynk\AiAttributes\Contracts\AIDriver;

public function boot(): void
{
    $this->app->make(AIManager::class)->extend('mistral', function ($app) {
        return new MistralDriver(config('ai-attributes.drivers.mistral'));
    });
}

Your driver only needs to implement one method:

class MistralDriver implements AIDriver
{
    public function __construct(protected array $config) {}

    public function generate(string $prompt, array $context = []): string
    {
        // Use Laravel's Http facade — the package itself does this for Claude/OpenAI.
        $response = Http::withToken($this->config['api_key'])
            ->post($this->config['base_url'].'/chat/completions', [
                'model' => $this->config['model'],
                'messages' => [['role' => 'user', 'content' => $prompt]],
            ]);

        return $response->json('choices.0.message.content');
    }
}

Configuration

The published config/ai-attributes.php is fully commented. Highlights:

return [
    'default' => env('AI_ATTRIBUTES_DRIVER', 'claude'),

    'cache' => [
        'enabled' => env('AI_ATTRIBUTES_CACHE_ENABLED', true),
        'store'   => env('AI_ATTRIBUTES_CACHE_STORE'),
        'ttl'     => (int) env('AI_ATTRIBUTES_CACHE_TTL', 60 * 60 * 24 * 30),
        'prefix'  => env('AI_ATTRIBUTES_CACHE_PREFIX', 'ai_attr'),
    ],

    'drivers' => [
        'claude' => [ /* api_key, base_url, model, max_tokens, timeout, version */ ],
        'openai' => [ /* api_key, base_url, model, max_tokens, timeout */ ],
        'ollama' => [ /* base_url, model, temperature, timeout */ ],
    ],
];

Testing

composer install
composer test

Tests use Pest and Orchestra Testbench. HTTP calls are faked with Http::fake() so the test suite never touches a real provider.

Roadmap

Shipped:

• Phase 1 — Core trait, Claude + OpenAI drivers, content-hash caching.
• Phase 2 — Per-attribute config, format casting (text/json/number/bool), retries with backoff, queued generation, runtime persona override, Artisan regenerate command.
• Phase 3 — Ollama driver ✨ Local LLMs via Ollama with temperature control and configurable base URL (works with remote Ollama servers too).

Coming:

• Phase 4 — Filament admin UI integration (a paid companion plugin).
• Phase 5 — Optional DB persistence, events, token-usage tracking, additional drivers (Gemini, Groq, OpenRouter), streaming, embeddings + RAG.

Contributing

Issues and PRs welcome. Run the test suite (composer test) and the formatter (composer format) before submitting.

Credits

• Reza
• All contributors

License

The MIT License (MIT). See LICENSE.md.