Skip to main content

​
Overview

LLM Magic provides multiple extraction strategies to handle different document sizes and complexity levels. Each strategy balances speed, cost, and accuracy differently.

​
Available Strategies

Access the list of available strategies: 
use Mateffy\Magic;

$strategies = Magic::getExtractionStrategies();
// Returns: Collection<string, class-string<Strategy>>
Built-in strategies:
  • simple - Single-pass extraction
  • sequential - Sequential batch processing
  • sequential-auto-merge - Sequential with automatic merging
  • parallel - Parallel batch processing
  • parallel-auto-merge - Parallel with automatic merging
  • double-pass - Two-pass extraction for maximum accuracy
  • double-pass-auto-merge - Two-pass with automatic merging

​
Simple Strategy

The simplest and fastest strategy. Processes only the first batch of artifacts. 
use Mateffy\Magic;

$data = Magic::extract()
    ->schema([
        'type' => 'object',
        'properties' => [
            'title' => ['type' => 'string'],
            'summary' => ['type' => 'string'],
        ]
    ])
    ->artifacts([$document])
    ->strategy('simple')
    ->send();
Best for: Short documents, quick summaries, single-page PDFsPros: Fastest, lowest costCons: Only processes first batch of artifacts

​
Sequential Strategy

Processes artifacts in batches sequentially, passing previous data to the next batch for context. 
$data = Magic::extract()
    ->schema($schema)
    ->artifacts($documents)
    ->strategy('sequential')
    ->send();
Best for: Long documents, maintaining context across pagesPros: Maintains context, processes all artifactsCons: Slower than parallel, sequential processing

​
Parallel Strategy

Processes batches concurrently, then merges results using an LLM. 
$data = Magic::extract()
    ->schema($schema)
    ->artifacts($documents)
    ->strategy('parallel')
    ->concurrency(5)  // Process 5 batches at once
    ->send();
Best for: Large documents where speed is criticalPros: Fastest for large documents, processes all artifactsCons: Higher cost, requires merge step, may lose some context

​
Double-Pass Strategy

Performs two passes: first parallel for broad coverage, then sequential for detail and accuracy. 
$data = Magic::extract()
    ->schema($schema)
    ->artifacts($documents)
    ->strategy('double-pass')
    ->send();
Best for: Complex documents requiring high accuracyPros: Highest accuracy, catches details missed in first passCons: Highest cost, slowest processing time

​
Auto-Merge Variants

Strategies ending in -auto-merge skip the LLM merge step and use automatic data merging: 
// Uses SmartDataMerger instead of LLM merge
->strategy('parallel-auto-merge')
Auto-merge variants are faster and cheaper but may produce less coherent results for complex schemas.

​
Registering Custom Strategies

Extend the extraction system with custom strategies: 
use Mateffy\Magic;
use App\Extraction\MyCustomStrategy;

Magic::registerStrategy('my-custom', MyCustomStrategy::class);

// Use it
Magic::extract()
    ->strategy('my-custom')
    ->send();
Your custom strategy must extend Extractor: 
use Mateffy\Magic\Extraction\Strategies\Extractor;

class MyCustomStrategy extends Extractor
{
    public function run(array $artifacts): array
    {
        // Your implementation
    }

    public static function getLabel(): string
    {
        return 'My Custom Strategy';
    }

    public function getEstimatedSteps(array $artifacts): int
    {
        return 1;
    }
}

​
Strategy Comparison

Simple

Speed: FastestCost: LowestAccuracy: BasicUse case: Short documents

Sequential

Speed: ModerateCost: ModerateAccuracy: GoodUse case: Long documents

Parallel

Speed: FastCost: HigherAccuracy: GoodUse case: Large documents

Double-Pass

Speed: SlowestCost: HighestAccuracy: BestUse case: Complex extraction

​
Performance Tips

Adjust chunk size based on your model’s context window:
->chunkSize(100000)  // For larger context models
Control how many batches process simultaneously:
->concurrency(10)  // Process up to 10 batches at once
Track extraction progress with callbacks:
->onDataProgress(function (array $data) {
    Log::info('Extraction progress', ['data' => $data]);
})

Build docs developers (and LLMs) love

Get started for freeTalk to us