EmbedVector Laravel Package

A Laravel package for building intelligent recommendation systems using OpenAI embeddings. Perfect for creating personalized content recommendations, job matching, product suggestions, and similar features where you need to find relevant matches based on user profiles or content similarity.

Features

Batch embedding processing using OpenAI's batch API
Separate database connection support for vector operations
Automatic vector extension creation for PostgreSQL
Efficient batch processing with configurable chunk sizes
Dual Contract System: Separate contracts for embedding generation and searchable models
Smart Model Separation: Models can be either embedding sources or searchable targets

Installation

Install the package via Composer:

composer require thesubhendu/embedvector-laravel

Publish the configuration and migrations:

php artisan vendor:publish --provider="Subhendu\EmbedVector\EmbedVectorServiceProvider"

Configure your environment variables:

OPENAI_API_KEY=your_openai_api_key_here

Database Requirements: This package requires PostgreSQL with the pgvector extension for vector operations.

Optional: If you want to use a separate PostgreSQL database connection other than your application database for vector operations, you can set the EMBEDVECTOR_DB_CONNECTION environment variable.

EMBEDVECTOR_DB_CONNECTION=pgsql

Run the migrations

php artisan migrate

Usage

Understanding the Contract System

This package uses two distinct contracts to separate concerns based on the direction of matching:

EmbeddableContract - For models that generate embeddings (e.g., Customer/Candidate profiles)
EmbeddingSearchableContract - For models that can be found using embeddings (e.g., Jobs)

Example Use Case: Job Matching for Candidates

If system is designed to find matching jobs for customers/candidates, not the other way around:

Customer/Candidate implements EmbeddableContract → generates embeddings from their profile, skills, preferences
Job implements EmbeddingSearchableContract → can be found/recommended based on candidate embeddings
Flow: Customer embeddings are used to find relevant Jobs that match their profile

For Bidirectional Matching: If you want both ways (finding jobs for candidates AND finding candidates for jobs), then both models need to implement EmbeddingSearchableContract.

Basic Embedding

use Subhendu\EmbedVector\Services\EmbeddingService;

$embeddingService = app(EmbeddingService::class);
$embedding = $embeddingService->createEmbedding('Your text here');

Implementing Contracts

For Models That Generate Embeddings (e.g., Customer)

use Illuminate\Database\Eloquent\Model;
use Subhendu\EmbedVector\Contracts\EmbeddableContract;
use Subhendu\EmbedVector\Traits\EmbeddableTrait;

class Customer extends Model implements EmbeddableContract
{
    use EmbeddableTrait;

    public function toEmbeddingText(): string
    {
        return $this->name . ' ' . $this->department . ' ' . $this->skills;
    }
}

For Models That Can Be Searched (e.g., Job)

use Illuminate\Database\Eloquent\Model;
use Illuminate\Database\Eloquent\Factories\HasFactory;
use Subhendu\EmbedVector\Contracts\EmbeddingSearchableContract;
use Subhendu\EmbedVector\Traits\EmbeddingSearchableTrait;

class Job extends Model implements EmbeddingSearchableContract
{
    use EmbeddingSearchableTrait;
    use HasFactory;

    public function toEmbeddingText(): string
    {
        return $this->title . ' ' . $this->description . ' ' . $this->requirements;
    }
}

Note: EmbeddingSearchableContract extends EmbeddableContract, and EmbeddingSearchableTrait automatically includes EmbeddableTrait functionality, so you only need to use one trait.

Finding Matching Results

Basic Usage

// Find jobs that match a customer's profile
$customer = Customer::find(1);
$matchingJobs = $customer->matchingResults(Job::class, 10);

foreach ($matchingJobs as $job) {
    echo "Job: {$job->title} - Match: {$job->match_percent}%";
    echo "Distance: {$job->distance}";
}

Note: The matchingResults() method automatically uses getOrCreateEmbedding() internally, which means:

If no embedding exists for the source model, it will be created
If an embedding exists but needs sync (embedding_sync_required = true), it will be updated
This ensures you always get accurate similarity results

Advanced Usage with Filters

You can add query filters to narrow down the search results before embedding similarity is calculated:

// Find only active jobs in specific locations
$customer = Customer::find(1);
$matchingJobs = $customer->matchingResults(
    targetModelClass: Job::class,
    topK: 10,
    queryFilter: function ($query) {
        $query->where('status', 'active')
              ->whereIn('location', ['New York', 'San Francisco'])
              ->where('salary', '>=', 80000);
    }
);

Method Parameters

targetModelClass (string): The class name of the model you want to find matches for
topK (int, default: 5): Maximum number of results to return
queryFilter (Closure, optional): Custom query constraints to apply before similarity matching

Return Properties

Each returned model includes additional properties:

match_percent (float): Similarity percentage (0-100, higher is better)
distance (float): Vector distance (lower is better for similarity)

Configuration

The package publishes a configuration file to config/embedvector.php with the following options:

return [
    'openai_api_key' => env('OPENAI_API_KEY', ''),
    'embedding_model' => env('EMBEDVECTOR_MODEL', 'text-embedding-3-small'),
    'distance_metric' => env('EMBEDVECTOR_DISTANCE', 'cosine'), // cosine | l2
    'search_strategy' => env('EMBEDVECTOR_SEARCH_STRATEGY', 'auto'), // auto | optimized | cross_connection
    'lot_size' => env('EMBEDVECTOR_LOT_SIZE', 50000),
    'chunk_size' => env('EMBEDVECTOR_CHUNK_SIZE', 500),
    'directories' => [
        'input' => 'embeddings/input',
        'output' => 'embeddings/output',
    ],
    'database_connection' => env('EMBEDVECTOR_DB_CONNECTION', 'pgsql'),
    'model_fields_to_check' => [
        // Configure fields to monitor for automatic sync
        // 'App\Models\Job' => ['title', 'description', 'requirements'],
    ],
];

Configuration Options Explained

openai_api_key: Your OpenAI API key (required in production)
embedding_model: OpenAI embedding model to use (text-embedding-3-small, text-embedding-3-large, etc.)
distance_metric: Vector similarity calculation method
- cosine: Better for semantic similarity (recommended)
- l2: Euclidean distance for geometric similarity
search_strategy: How to perform similarity searches
- auto: Automatically choose the best strategy (recommended)
- optimized: Use JOIN-based queries (same database only)
- cross_connection: Two-step approach (works across different databases)
lot_size: Maximum items per OpenAI batch (up to 50,000)
chunk_size: Items processed per chunk during batch generation
database_connection: PostgreSQL connection for vector operations
model_fields_to_check: Configure fields to monitor for automatic sync with FireSyncEmbeddingTrait

Batch Processing

For processing large datasets efficiently, this package provides batch processing capabilities using OpenAI's batch API, which is more cost-effective for processing many embeddings at once.

Commands

php artisan embedding:gen {model} {--type=sync|init} {--force} - Generate batch embeddings for a specific model
php artisan embedding:proc {--batch-id=} {--all} - Process completed batch results

Command Options

`embedding:gen`

{model} - The model class name to generate embeddings for (e.g. App\\Models\\Job)
--type=sync - Processing type (default: sync)
--force - Force overwrite existing files

`embedding:proc`

--batch-id= - Process a specific batch by ID
--all - Process all completed batches
No options - Check and process batches that are ready (default behavior)

Usage Examples

# Generate embeddings for User model (init = first time, sync = update existing)
php artisan embedding:gen "App\\Models\\User" --type=init

# Generate embeddings for sync (only models that need updates)
php artisan embedding:gen "App\\Models\\Job" --type=sync

# Check and process ready batches (default)
php artisan embedding:proc

# Process all completed batches
php artisan embedding:proc --all

# Process specific batch
php artisan embedding:proc --batch-id=batch_abc123

Real-World Examples

E-commerce Product Recommendations

// Product model (searchable)
class Product extends Model implements EmbeddingSearchableContract
{
    use EmbeddingSearchableTrait;

    public function toEmbeddingText(): string
    {
        return $this->name . ' ' . $this->description . ' ' . $this->category . ' ' . $this->tags;
    }
}

// User model (generates embeddings from purchase history)
class User extends Model implements EmbeddableContract
{
    use EmbeddableTrait;

    public function toEmbeddingText(): string
    {
        $purchaseHistory = $this->orders()
            ->with('products')
            ->get()
            ->flatMap->products
            ->pluck('name')
            ->implode(' ');
            
        return $this->preferences . ' ' . $purchaseHistory;
    }
}

// Find recommended products for a user
$user = User::find(1);
$recommendations = $user->matchingResults(
    targetModelClass: Product::class,
    topK: 20,
    queryFilter: function ($query) {
        $query->where('in_stock', true)
              ->where('price', '<=', 500)
              ->whereNotIn('id', auth()->user()->purchased_product_ids);
    }
);

Job Matching Platform

// Find jobs for a candidate with filters
$candidate = Candidate::find(1);
$matchingJobs = $candidate->matchingResults(
    targetModelClass: Job::class,
    topK: 15,
    queryFilter: function ($query) use ($candidate) {
        $query->where('status', 'open')
              ->where('remote_allowed', $candidate->prefers_remote)
              ->whereIn('experience_level', $candidate->acceptable_levels)
              ->where('salary_min', '>=', $candidate->min_salary);
    }
);

foreach ($matchingJobs as $job) {
    echo "Match: {$job->match_percent}% - {$job->title} at {$job->company}";
}

Content Recommendation System

// Article model
class Article extends Model implements EmbeddingSearchableContract
{
    use EmbeddingSearchableTrait;

    public function toEmbeddingText(): string
    {
        return $this->title . ' ' . $this->summary . ' ' . $this->tags . ' ' . $this->category;
    }
}

// User reading history model
class UserProfile extends Model implements EmbeddableContract
{
    use EmbeddableTrait;

    public function toEmbeddingText(): string
    {
        $readingHistory = $this->user->readArticles()
            ->selectRaw('GROUP_CONCAT(title, " ", summary) as content')
            ->value('content');
            
        return $this->interests . ' ' . $readingHistory;
    }
}

// Get personalized article recommendations
$profile = UserProfile::where('user_id', auth()->id())->first();
$recommendations = $profile->matchingResults(
    targetModelClass: Article::class,
    topK: 10,
    queryFilter: function ($query) use ($profile) {
        $query->where('published', true)
              ->where('created_at', '>=', now()->subDays(7))
              ->whereNotIn('id', $profile->user->read_article_ids);
    }
);

Embedding Management Examples

Working with Embeddings

$job = Job::find(1);

// Check if an embedding exists without creating one
$embedding = $job->getEmbedding();
if ($embedding) {
    echo "Embedding exists: " . ($embedding->embedding_sync_required ? "Needs sync" : "Up to date");
} else {
    echo "No embedding found";
}

// Get or create embedding (will create if missing or update if sync required)
$embedding = $job->getOrCreateEmbedding();
echo "Embedding ready with match percentage calculation";

// Force create a fresh embedding (useful for testing or manual refresh)
$freshEmbedding = $job->createFreshEmbedding();

// Queue for syncing (mark for batch update later)
$job->queueForSyncing();

Batch Sync Workflow

// 1. Mark multiple models for syncing
$jobs = Job::where('updated_at', '>', now()->subDays(1))->get();
foreach ($jobs as $job) {
    $job->queueForSyncing(); // Queue each job for sync
}

// 2. Process all queued embeddings in batch
php artisan embedding:gen "App\\Models\\Job" --type=sync

// 3. Process the completed batch
php artisan embedding:proc --all

Conditional Embedding Updates

class JobController extends Controller
{
    public function update(Request $request, Job $job)
    {
        $job->update($request->validated());
        
        // Only queue for syncing if embedding-relevant fields changed
        if ($job->wasChanged(['title', 'description', 'requirements'])) {
            $job->queueForSyncing();
        }
        
        return response()->json($job);
    }
}

Real-time vs Batch Embedding Strategy

// Real-time embedding (immediate, good for single updates)
$job = Job::create($data);
$embedding = $job->getOrCreateEmbedding(); // Creates immediately

// Batch embedding (efficient for bulk updates)
$jobs = Job::factory()->count(100)->create();
foreach ($jobs as $job) {
    $job->queueForSyncing(); // Mark for batch processing
}
// Then run: php artisan embedding:gen "App\\Models\\Job" --type=sync

Embedding Lifecycle Management

// Scenario 1: New model creation
$job = Job::create($data);
// Option A: Create embedding immediately
$embedding = $job->getOrCreateEmbedding();
// Option B: Queue for batch processing (more efficient)
$job->queueForSyncing();

// Scenario 2: Model updates
$job->update(['title' => 'Updated Title']);
// Option A: Update embedding immediately
$job->createFreshEmbedding();
// Option B: Queue for batch processing (recommended)
$job->queueForSyncing();

// Scenario 3: Checking embedding status
$embedding = $job->getEmbedding();
if (!$embedding) {
    echo "No embedding exists";
} elseif ($embedding->embedding_sync_required) {
    echo "Embedding needs update";
} else {
    echo "Embedding is up to date";
}

// Scenario 4: Bulk operations
$jobs = Job::where('department', 'Engineering')->get();
foreach ($jobs as $job) {
    $job->queueForSyncing(); // Queue all for batch processing
}
// Process in batch: php artisan embedding:gen "App\\Models\\Job" --type=sync

Best Practices

1. Optimize Your `toEmbeddingText()` Method

public function toEmbeddingText(): string
{
    // ✅ Good: Concise, relevant information
    return trim($this->title . ' ' . $this->description . ' ' . $this->tags);
    
    // ❌ Avoid: Too much noise or irrelevant data
    // return $this->created_at . ' ' . $this->id . ' ' . $this->long_legal_text;
}

2. Use Appropriate Filters

// ✅ Good: Filter before similarity calculation
$matches = $user->matchingResults(
    Product::class,
    10,
    fn($q) => $q->where('available', true)->where('price', '<=', $budget)
);

// ❌ Less efficient: Filtering after embedding calculation
$allMatches = $user->matchingResults(Product::class, 100);
$filtered = $allMatches->where('available', true);

3. Choose the Right Embedding Method

Understanding when to use each embedding method:

$job = Job::find(1);

// ✅ Use getEmbedding() when you just want to check if embedding exists
$embedding = $job->getEmbedding();
if ($embedding && !$embedding->embedding_sync_required) {
    // Use existing embedding
}

// ✅ Use getOrCreateEmbedding() for similarity matching (recommended)
$matchingJobs = $customer->matchingResults(Job::class); // Uses getOrCreateEmbedding internally

// ✅ Use createFreshEmbedding() when you want to force regeneration
$job->update(['title' => 'New Title']);
$freshEmbedding = $job->createFreshEmbedding(); // Immediate update

// ✅ Use queueForSyncing() for deferred batch processing (most efficient)
$job->update(['title' => 'New Title']);
$job->queueForSyncing(); // Mark for later batch processing

4. Manage Embedding Sync

Manual Sync Management

// Method 1: Using queueForSyncing() (recommended)
class Job extends Model implements EmbeddingSearchableContract 
{
    use EmbeddingSearchableTrait;
    
    protected static function booted()
    {
        static::updated(function ($job) {
            if ($job->isDirty(['title', 'description', 'requirements'])) {
                $job->queueForSyncing(); // Simpler and cleaner approach
            }
        });
    }
}

// Method 2: Direct embedding update (legacy approach)
class Job extends Model implements EmbeddingSearchableContract 
{
    use EmbeddingSearchableTrait;
    
    protected static function booted()
    {
        static::updated(function ($job) {
            if ($job->isDirty(['title', 'description', 'requirements'])) {
                $job->embedding()->update(['embedding_sync_required' => true]);
            }
        });
    }
}

Automatic Sync Management with FireSyncEmbeddingTrait

For automatic embedding sync management, use the FireSyncEmbeddingTrait:

use Subhendu\EmbedVector\Traits\FireSyncEmbeddingTrait;

class Job extends Model implements EmbeddingSearchableContract 
{
    use EmbeddingSearchableTrait, FireSyncEmbeddingTrait;
    
    // No need for manual booted() method - trait handles it automatically
}

Configure which fields to monitor in your config/embedvector.php:

return [
    // ... other config options
    
    'model_fields_to_check' => [
        'App\Models\Job' => ['title', 'description', 'requirements'],
        'App\Models\Product' => ['name', 'description', 'category'],
        'App\Models\User' => ['name', 'bio', 'skills'],
    ],
];

How it works:

The trait automatically monitors specified fields for changes
When any monitored field changes, it marks the embedding for re-sync
Only triggers when fields actually change (compares old vs new values)
Respects the configuration mapping for each model class

Troubleshooting

Common Issues

"No embedding found in response"
- Check your OpenAI API key is valid
- Verify the embedding model exists
- Ensure your toEmbeddingText() returns non-empty strings
"Model class must implement EmbeddingSearchableContract"
- Target models must implement EmbeddingSearchableContract
- Source models only need EmbeddableContract
Poor matching results
- Review your toEmbeddingText() method - it should contain relevant, semantic information
- Consider using cosine distance for semantic similarity
- Try different embedding models (text-embedding-3-large for better quality)
Performance issues
- Use batch processing for large datasets
- Consider using optimized search strategy for same-database scenarios
- Add appropriate database indexes
Cross-connection relationship limitations
- The embedding() relationship only works when both models use the same database connection
- For cross-connection setups (e.g., Jobs in MySQL, embeddings in PostgreSQL), use getEmbedding() or getOrCreateEmbedding() methods instead of the relationship
- Direct relationship access ($model->embedding) will return null in cross-connection scenarios
Embedding method confusion
- Use getEmbedding() when you only want to check if an embedding exists (returns null if not found)
- Use getOrCreateEmbedding() when you need an embedding for similarity matching (creates/updates as needed)
- Use queueForSyncing() to defer embedding updates for batch processing (most efficient for bulk updates)

Database Performance

-- Add indexes for better performance
CREATE INDEX IF NOT EXISTS embeddings_model_type_idx ON embeddings (model_type);
CREATE INDEX IF NOT EXISTS embeddings_sync_required_idx ON embeddings (embedding_sync_required);

Environment Variables Reference

Add these to your .env file:

# Required
OPENAI_API_KEY=your_openai_api_key_here

# Optional - Customize behavior
EMBEDVECTOR_MODEL=text-embedding-3-small
EMBEDVECTOR_DISTANCE=cosine
EMBEDVECTOR_SEARCH_STRATEGY=auto
EMBEDVECTOR_LOT_SIZE=50000
EMBEDVECTOR_CHUNK_SIZE=500
EMBEDVECTOR_DB_CONNECTION=pgsql

Architecture Overview

┌─────────────────┐    ┌─────────────────┐    ┌─────────────────┐
│   Source Model  │    │  Target Model   │    │   Embeddings    │
│ (EmbeddableContract)  │ (EmbeddingSearchableContract) │    │    Table       │
│                 │    │                 │    │                 │
│ • Customer      │───▶│ • Job           │◀──│ • Vector data   │
│ • User Profile  │    │ • Product       │   │ • Similarity    │
│ • Candidate     │    │ • Article       │   │   calculations  │
└─────────────────┘    └─────────────────┘    └─────────────────┘
         │                       │                       │
         ▼                       ▼                       ▼
┌─────────────────┐    ┌─────────────────┐    ┌─────────────────┐
│ toEmbeddingText()│    │ toEmbeddingText()│    │  PostgreSQL     │
│ • Generate text │    │ • Generate text │    │  with pgvector  │
│   for embedding │    │   for embedding │    │  extension      │
└─────────────────┘    └─────────────────┘    └─────────────────┘

API Reference

EmbeddableTrait Methods

`matchingResults(string $targetModelClass, int $topK = 5, ?Closure $queryFilter = null): Collection`

Find models similar to the current model.

Parameters:

$targetModelClass: Fully qualified class name of the target model
$topK: Maximum number of results (default: 5)
$queryFilter: Optional closure to filter results before similarity calculation

Returns: Collection of models with match_percent and distance properties

`getEmbedding(): ?Embedding`

Get the existing embedding for the current model without creating a new one.

Returns: Embedding model instance or null if no embedding exists

`getOrCreateEmbedding(): Embedding`

Get the existing embedding or create a new one if none exists. Also handles updating embeddings when embedding_sync_required is true.

Returns: Embedding model instance

`queueForSyncing(): void`

Mark the model's embedding for re-generation on the next sync. This is useful when you want to defer embedding updates until a batch process runs.

Returns: void

`createFreshEmbedding(): Embedding`

Force create a new embedding for the model, bypassing any existing embedding.

Returns: Newly created Embedding model instance

`embedding(): MorphOne`

Eloquent relationship to the embedding record.

Returns: MorphOne relationship

EmbeddingSearchableTrait Methods

`queryForEmbedding(): Builder`

Get the base query for models to be embedded during initial processing.

Returns: Eloquent Builder instance

`queryForSyncing(): Builder`

Get the query for models that need re-embedding (sync process).

Returns: Eloquent Builder instance

Configuration Methods

`getConnectionName(): ?string`

Get the database connection name for the model.

Returns: Database connection name or null for default

Testing

The package includes comprehensive tests. Run them with:

# Run all tests
vendor/bin/pest

# Run with coverage (requires Xdebug)
vendor/bin/pest --coverage

# Run static analysis
vendor/bin/phpstan analyse

Contributing

Fork the repository
Create your feature branch (git checkout -b feature/amazing-feature)
Make your changes
Add tests for your changes
Ensure all tests pass (vendor/bin/pest)
Run static analysis (vendor/bin/phpstan analyse)
Commit your changes (git commit -m 'Add amazing feature')
Push to the branch (git push origin feature/amazing-feature)
Open a Pull Request

License

The MIT License (MIT). Please see License File for more information.

Credits

Subhendu Bhatta
Built with Laravel
Powered by OpenAI Embeddings
Uses pgvector for PostgreSQL vector operations

Name		Name	Last commit message	Last commit date
Latest commit History 81 Commits
config		config
database		database
resources/views		resources/views
src		src
tests		tests
workbench/app/Providers		workbench/app/Providers
.editorconfig		.editorconfig
.gitattributes		.gitattributes
.gitignore		.gitignore
CHANGELOG.md		CHANGELOG.md
LICENSE.md		LICENSE.md
README.md		README.md
composer.json		composer.json
phpstan-baseline.neon		phpstan-baseline.neon
phpstan.neon.dist		phpstan.neon.dist
phpunit.xml.dist		phpunit.xml.dist

License

thesubhendu/embedvector-laravel

Folders and files

Latest commit

History

Repository files navigation

EmbedVector Laravel Package

Features

Installation

Usage

Understanding the Contract System

Example Use Case: Job Matching for Candidates

Basic Embedding

Implementing Contracts

For Models That Generate Embeddings (e.g., Customer)

For Models That Can Be Searched (e.g., Job)

Finding Matching Results

Basic Usage

Advanced Usage with Filters

Method Parameters

Return Properties

Configuration

Configuration Options Explained

Batch Processing

Commands

Command Options

embedding:gen

embedding:proc

Usage Examples

Real-World Examples

E-commerce Product Recommendations

Job Matching Platform

Content Recommendation System

Embedding Management Examples

Working with Embeddings

Batch Sync Workflow

Conditional Embedding Updates

Real-time vs Batch Embedding Strategy

Embedding Lifecycle Management

Best Practices

1. Optimize Your toEmbeddingText() Method

2. Use Appropriate Filters

3. Choose the Right Embedding Method

4. Manage Embedding Sync

Manual Sync Management

Automatic Sync Management with FireSyncEmbeddingTrait

Troubleshooting

Common Issues

Database Performance

Environment Variables Reference

Architecture Overview

API Reference

EmbeddableTrait Methods

matchingResults(string $targetModelClass, int $topK = 5, ?Closure $queryFilter = null): Collection

getEmbedding(): ?Embedding

getOrCreateEmbedding(): Embedding

queueForSyncing(): void

createFreshEmbedding(): Embedding

embedding(): MorphOne

EmbeddingSearchableTrait Methods

queryForEmbedding(): Builder

queryForSyncing(): Builder

Configuration Methods

getConnectionName(): ?string

Testing

Contributing

License

Credits

About

Resources

License

Uh oh!

Stars

Watchers

Forks

Releases

Packages 0

Uh oh!

Contributors 4

Uh oh!

`embedding:gen`

`embedding:proc`

1. Optimize Your `toEmbeddingText()` Method

`matchingResults(string $targetModelClass, int $topK = 5, ?Closure $queryFilter = null): Collection`

`getEmbedding(): ?Embedding`

`getOrCreateEmbedding(): Embedding`

`queueForSyncing(): void`

`createFreshEmbedding(): Embedding`

`embedding(): MorphOne`

`queryForEmbedding(): Builder`

`queryForSyncing(): Builder`

`getConnectionName(): ?string`

Packages