Laravel Prompt Manager Laravel Package

Getting Started

Minimal Steps

1. Installation

   composer require prismaticoder/laravel-prompt-manager
   php artisan vendor:publish --provider="Prismaticoder\PromptManager\PromptManagerServiceProvider" --tag="migrations"
   php artisan migrate
   

2. Define a Prompt Create a prompt template in resources/views/prompts/your-prompt.blade.php:

   {{ $userInput }}
   {{ $systemInstructions }}
   

3. First Use Case Register and use the prompt in a controller:

   use Prismaticoder\PromptManager\Facades\PromptManager;
   
   public function generateResponse()
   {
       $prompt = PromptManager::prompt('your-prompt')
           ->withVariables([
               'userInput' => 'Hello AI',
               'systemInstructions' => 'Be concise.'
           ])
           ->get();
   
       return $prompt->execute();
   }
   

Where to Look First

• Migrations: Check database/migrations/ for the prompt_versions table structure.
• Config: config/prompt-manager.php for default settings (e.g., default_model, token_estimation).
• Facade: app/Facades/PromptManager.php for quick usage examples.

Implementation Patterns

Core Workflows

1. Prompt Definition & Versioning

   // Define a new prompt (auto-versions on save)
   PromptManager::prompt('greeting')
       ->setTemplate('resources/views/prompts/greeting.blade.php')
       ->setDescription('Handles user greetings')
       ->save();
   
   // Update a prompt (creates a new version)
   PromptManager::prompt('greeting')
       ->setTemplate('updated-template.blade.php')
       ->save();
   

2. Dynamic Variable Injection

   $response = PromptManager::prompt('analysis')
       ->withVariables([
           'data' => $userData,
           'tone' => 'professional',
       ])
       ->execute(); // Uses configured LLM (e.g., OpenAI)
   

3. Model-Specific Prompts

   // Tag a prompt for a specific model
   PromptManager::prompt('gpt4-only')
       ->addModelTag('gpt-4')
       ->save();
   
   // Select version based on model context
   $version = PromptManager::prompt('gpt4-only')
       ->forModel('gpt-4')
       ->get();
   

4. A/B Testing

   // Register variants
   PromptManager::prompt('ab-test')
       ->addVariant('variant_a', 'template-a.blade.php')
       ->addVariant('variant_b', 'template-b.blade.php')
       ->save();
   
   // Serve randomly or by condition
   $variant = PromptManager::prompt('ab-test')
       ->selectVariant('random')
       ->get();
   

Integration Tips

• Service Providers: Bind the PromptManager to your container for dependency injection:

  $this->app->bind('promptManager', function ($app) {
      return new \Prismaticoder\PromptManager\PromptManager(
          $app['config']['prompt-manager']
      );
  });
  

• Event Listeners: Listen for PromptSaved or PromptVersionCreated to log/notify:

  public function handle(PromptVersionCreated $event)
  {
      Log::info("New prompt version created: {$event->version->name}");
  }
  

• Testing: Use the PromptManager facade in tests to mock responses:

  PromptManager::shouldReceive('execute')
      ->once()
      ->andReturn('Mocked response');
  

Gotchas and Tips

Pitfalls

1. Template Paths

   • Issue: Blade templates must be in resources/views/prompts/ or the view_path config will fail.
   • Fix: Verify config/prompt-manager.php has the correct view_path.

2. Variable Validation

   • Issue: Undefined variables in templates cause silent failures (no error thrown).
   • Fix: Use ->withRequiredVariables(['key' => 'default']) to enforce presence.

3. Model Context Overrides

   • Issue: Forgetting to specify a model for tagged prompts may fall back to default_model, which could be incorrect.
   • Fix: Always use ->forModel('model-name') when model-specific logic exists.

4. Token Estimation

   • Issue: Custom token counting may misalign with LLM expectations (e.g., OpenAI vs. Anthropic).
   • Fix: Override the tokenEstimator in config or extend the TokenEstimator class.

Debugging

• Prompt Rendering: Use ->debug() to dump the rendered template:

  $prompt = PromptManager::prompt('debug-me')->debug();
  

• Version History: Check the prompt_versions table directly or use:

  PromptManager::prompt('history')->versions()->get();
  

• LLM Errors: Wrap executions in try-catch to log LLM-specific errors (e.g., rate limits):

  try {
      $response = $prompt->execute();
  } catch (\Prismaticoder\PromptManager\Exceptions\LLMException $e) {
      Log::error($e->getMessage());
  }
  

Extension Points

1. Custom Storage Override the PromptRepository to use a different storage backend (e.g., Redis):

   $promptRepo = new \Prismaticoder\PromptManager\Repositories\RedisPromptRepository();
   PromptManager::setRepository($promptRepo);
   

2. Prompt Validators Add validation logic before execution:

   PromptManager::prompt('secure')
       ->addValidator(function ($prompt) {
           return strlen($prompt->template) < 1000; // Reject long prompts
       });
   

3. Event Extensions Dispatch custom events for prompt lifecycle hooks:

   // In PromptManagerServiceProvider
   Event::listen('prompt.before_execute', function ($prompt) {
       // Pre-execution logic
   });
   

4. Token Calculators Replace the default token estimator for custom models:

   $estimator = new \Prismaticoder\PromptManager\Services\CustomTokenEstimator();
   PromptManager::setTokenEstimator($estimator);
   

Config Quirks

• Default Model: Ensure config/prompt-manager.php has default_model set to your primary LLM (e.g., 'gpt-3.5-turbo').
• Caching: Disable cache_enabled in config if testing frequently (versions won’t persist).
• Environment Variables: Use .env for sensitive LLM API keys (e.g., OPENAI_API_KEY).