Cortex - Laravel LLM Package

Cortex is a Laravel 12 package for interacting with Large Language Models (LLMs). It features a robust plugin system for extensibility and provides a clean, fluent API for AI-powered applications.

Features

• Plugin-Based Architecture - Modular design with core and optional plugins
• Multiple Providers - AWS Bedrock support with more coming
• Streaming Support - Real-time streaming with Laravel Echo and SSE broadcasting
• Tool/Function Calling - Define and execute tools with automatic schema generation
• Structured Output - Enforce typed responses from LLMs with validation
• Schema Validation - JSON Schema generation and validation with casting
• Autonomous Agents - Build agents with memory, tools, and agentic loops
• MCP Integration - Connect to Model Context Protocol servers for extended capabilities
• Workflow Orchestration - Build complex multi-step processes with branching, loops, and human-in-the-loop
• Testing Support - FakeProvider for testing without API calls

Requirements

• PHP 8.2+
• Laravel 12.x
• AWS SDK (for Bedrock provider)

Installation

composer require jayi/cortex

Publish the configuration:

php artisan vendor:publish --tag=cortex-config

Quick Start

Basic Chat

use JayI\Cortex\Facades\Cortex;

// Simple message
$response = Cortex::chat()
    ->message('What is the capital of France?')
    ->send();

echo $response->content(); // "The capital of France is Paris."

Streaming

$stream = Cortex::chat()
    ->message('Write a short story about a robot.')
    ->stream();

foreach ($stream->text() as $chunk) {
    echo $chunk;
}

With Tools

use JayI\Cortex\Plugins\Tool\Tool;
use JayI\Cortex\Plugins\Schema\Schema;

$weatherTool = Tool::make('get_weather')
    ->withDescription('Get current weather for a location')
    ->withInput(
        Schema::object()
            ->property('location', Schema::string()->description('City name'))
            ->required('location')
    )
    ->withHandler(function (array $input) {
        return ToolResult::success(['temperature' => 22, 'conditions' => 'Sunny']);
    });

$response = Cortex::chat()
    ->message('What is the weather in Paris?')
    ->withTools([$weatherTool])
    ->send();

Structured Output

use JayI\Cortex\Plugins\Schema\Schema;

$schema = Schema::object()
    ->property('sentiment', Schema::enum(['positive', 'negative', 'neutral']))
    ->property('confidence', Schema::number()->minimum(0)->maximum(1))
    ->required('sentiment', 'confidence');

$response = Cortex::structuredOutput()->generate(
    Cortex::chat()->message('Analyze: "This product is amazing!"')->build(),
    $schema
);

// ['sentiment' => 'positive', 'confidence' => 0.95]

Autonomous Agents

use JayI\Cortex\Plugins\Agent\Agent;
use JayI\Cortex\Plugins\Agent\Memory\BufferMemory;

$agent = Agent::make('research-assistant')
    ->withSystemPrompt('You are a research assistant with access to web search.')
    ->withTools([$searchTool, $calculatorTool])
    ->withMemory(new BufferMemory())
    ->withMaxIterations(10);

$response = $agent->run('What is the population of Tokyo?');

echo $response->content;
echo "Completed in {$response->iterationCount} iterations";

Configuration

See config/cortex.php for all configuration options:

return [
    'plugins' => [
        'enabled' => [
            'tool',
            'structured-output',
        ],
    ],

    'provider' => [
        'default' => 'bedrock',
        'providers' => [
            'bedrock' => [
                'region' => env('AWS_REGION', 'us-east-1'),
                'default_model' => env('CORTEX_DEFAULT_MODEL', 'anthropic.claude-3-5-sonnet-20241022-v2:0'),
            ],
        ],
    ],
];

Documentation

• Plugin System
• Schema Plugin
• Provider Plugin
• Chat Plugin
• Tool Plugin
• Structured Output Plugin
• Agent Plugin
• MCP Plugin
• Workflow Plugin
• Resilience Plugin
• Usage Plugin
• Guardrail Plugin
• Cache Plugin
• Context Manager Plugin

Architecture

Cortex uses a plugin-based architecture where functionality is organized into discrete plugins:

Core Plugins (Always Loaded):

• Schema - JSON Schema generation and validation
• Provider - LLM provider management and abstraction
• Chat - Chat completion and streaming

Optional Plugins:

• Tool - Tool/function calling support
• Structured Output - Enforce typed responses
• Agent - Autonomous agents with memory and tool use
• MCP - Model Context Protocol integration
• Workflow - Multi-step workflow orchestration
• Resilience - Retry, circuit breaker, rate limiting, fallbacks
• Usage - Token tracking, cost estimation, budget management
• Guardrail - Content filtering, PII detection, prompt injection protection
• Cache - Response caching with exact and semantic matching
• Context Manager - Context window management and message pruning

Testing

Cortex provides a FakeProvider for testing:

use JayI\Cortex\Plugins\Provider\Providers\FakeProvider;

$fake = FakeProvider::text('Mocked response');

$fake->assertSentCount(0);

$response = $fake->chat($request);

$fake->assertSentCount(1);
$fake->assertSent(fn ($r) => str_contains($r->messages->last()->text(), 'expected'));

License

MIT License. See LICENSE for details.