thaolaptrinh/laravel-rag

Production-grade RAG engine for Laravel — provider-agnostic, HTTP-based, extensible

v0.1.0 2026-03-29 11:24 UTC

Requires

• php: ^8.4
• illuminate/console: ^11.0||^12.0||^13.0
• illuminate/contracts: ^11.0||^12.0||^13.0
• illuminate/database: ^11.0||^12.0||^13.0
• illuminate/http: ^11.0||^12.0||^13.0
• illuminate/log: ^11.0||^12.0||^13.0
• illuminate/support: ^11.0||^12.0||^13.0
• spatie/laravel-package-tools: ^1.16

Requires (Dev)

• larastan/larastan: ^3.0
• laravel/pint: ^1.14
• nunomaduro/collision: ^8.8
• orchestra/testbench: ^10.0.0||^9.0.0
• pestphp/pest: ^4.0
• pestphp/pest-plugin-arch: ^4.0
• pestphp/pest-plugin-laravel: ^4.0
• phpstan/extension-installer: ^1.4
• phpstan/phpstan-deprecation-rules: ^2.0
• phpstan/phpstan-phpunit: ^2.0
• phpstan/phpstan-strict-rules: ^2.0
• rector/rector: ^2.3

MIT

• Thao Nguyen Van

• laravel
• llm
• embeddings
• rag
• pgvector
• retrieval-augmented-generation
• thaolaptrinh

README

Production-grade RAG (Retrieval-Augmented Generation) engine for Laravel — provider-agnostic, HTTP-based, extensible.

Features

• Document Ingestion: Split documents into chunks, generate embeddings, store in pgvector
• Semantic Search: Cosine similarity search with metadata filtering
• LLM Integration: OpenAI-compatible API support (OpenAI, FPT Cloud, Ollama, vLLM, Supabase)
• Idempotent Re-ingestion: Skip unchanged documents using content hash
• Queue Support: Process large document sets asynchronously
• Testing Helpers: Fake mode for unit testing

Requirements

• PHP 8.4+
• Laravel 11.0, 12.0, 13.0
• PostgreSQL 13+ with pgvector extension

Installation

composer require thaolaptrinh/laravel-rag

Configuration

1. Add the RAG database connection to config/database.php:

'rag' => [
    'driver' => 'pgsql',
    'host' => env('RAG_DB_HOST', '127.0.0.1'),
    'port' => env('RAG_DB_PORT', '5432'),
    'database' => env('RAG_DB_DATABASE', 'laravel'),
    'username' => env('RAG_DB_USERNAME', 'postgres'),
    'password' => env('RAG_DB_PASSWORD', ''),
    'prefix' => '',
    'search_path' => env('RAG_DB_SCHEMA', 'public'),
],

1. Publish and run migrations:

php artisan rag:install

Or manually:

php artisan vendor:publish --provider="Thaolaptrinh\Rag\RagServiceProvider" --tag="rag-config"
php artisan vendor:publish --provider="Thaolaptrinh\Rag\RagServiceProvider" --tag="rag-migrations"
php artisan migrate

Post-Installation: Create HNSW Index

After ingesting your first documents, create the HNSW vector index for optimal search performance:

php artisan rag:ingest "Your document content" --id="doc-1"
php artisan rag:index

Note: The HNSW index cannot be created on an empty table. Run rag:index after your first ingestion.

Environment Variables

# Database (separate from app DB)
RAG_DB_CONNECTION=rag
RAG_DB_HOST=127.0.0.1
RAG_DB_PORT=5432
RAG_DB_DATABASE=laravel
RAG_DB_USERNAME=postgres
RAG_DB_PASSWORD=

# Embedding (OpenAI-compatible HTTP)
RAG_EMBEDDING_API_URL=https://api.openai.com/v1/embeddings
RAG_EMBEDDING_API_KEY=sk-...
RAG_EMBEDDING_MODEL=text-embedding-3-small
RAG_EMBEDDING_DIMENSIONS=1536
RAG_EMBEDDING_BATCH_SIZE=100
RAG_EMBEDDING_TIMEOUT=120

# LLM (OpenAI-compatible HTTP)
RAG_LLM_API_URL=https://api.openai.com/v1/chat/completions
RAG_LLM_API_KEY=sk-...
RAG_LLM_MODEL=gpt-4o-mini
RAG_LLM_MAX_OUTPUT_TOKENS=4096
RAG_LLM_CONTEXT_WINDOW=128000
RAG_LLM_TEMPERATURE=0.7
RAG_LLM_TIMEOUT=120

# Chunking
RAG_CHUNK_SIZE=1000
RAG_CHUNK_OVERLAP=200

# Document
RAG_DOCUMENT_MAX_CONTENT_LENGTH=100000

# Ingestion
RAG_INGESTION_SUB_BATCH_SIZE=10
RAG_INGESTION_PIPELINE_TIMEOUT=600

# Retrieval
RAG_RETRIEVAL_TOP_K=20
RAG_RETRIEVAL_MIN_SCORE=0.0
RAG_HNSW_EF_SEARCH=100

# Prompt
RAG_PROMPT_SYSTEM=You are a helpful assistant...
RAG_PROMPT_TOKENS_PER_CHAR=0.25

# Contextual Retrieval (Optional)
RAG_CONTEXTUAL_ENABLED=false

Usage

Ingest Documents

use Thaolaptrinh\Rag\Data\Document;
use Thaolaptrinh\Rag\Rag;

// Single document
$result = Rag::ingest(Document::create('Your document content here', [
    'source' => 'manual',
    'file_id' => 123,
]));

// Multiple documents
$result = Rag::ingestMany([
    Document::create('Content 1', ['source' => 'pdf']),
    Document::create('Content 2', ['source' => 'pdf']),
]);

Query

// Simple query
$answer = Rag::query('What is this document about?');

// With options
$answer = Rag::query('What is this about?', [
    'top_k' => 10,
    'filters' => ['source' => 'pdf'],
    'system' => 'Custom system prompt',
]);

echo $answer->text;
echo $answer->traceId;

// Access sources
foreach ($answer->sources as $source) {
    echo $source->content;
    echo $source->score;
}

Streaming

Rag::queryStream('Tell me about...', function (string $token): void {
    echo $token;
});

Delete

// Delete single document
Rag::delete('document-id');

// Delete multiple
$count = Rag::deleteMany(['id-1', 'id-2']);

// Delete all
Rag::truncate();

Queue Support

For large document sets, use queued ingestion:

// Queue single document
Rag::ingestQueued($document);
Rag::ingestQueued($document, 'rag-embeddings');

// Queue multiple (one job per document)
Rag::ingestManyQueued($documents);

Testing

use Thaolaptrinh\Rag\Rag;
use Thaolaptrinh\Rag\Data\Document;

beforeEach(function () {
    Rag::fake();
});

it('ingests documents', function () {
    $result = Rag::ingest(Document::create('test content'));
    
    expect($result->ingested)->toBe(1);
    
    Rag::assertIngested('document-id');
});

it('queries the store', function () {
    $answer = Rag::query('What is Laravel?');
    
    expect($answer->text)->toBe('Fake response');
    
    Rag::assertQueried('What is Laravel?');
});

Artisan Commands

# Ingest a document
php artisan rag:ingest "Document content" --id="doc-1" --metadata='{"source": "manual"}'

# Query
php artisan rag:query "What is this about?"

# Delete
php artisan rag:delete doc-1
php artisan rag:delete --all

# Install (publish config + run migrations)
php artisan rag:install
php artisan rag:install --without-migration

# Create HNSW index (after ingesting data)
php artisan rag:index

Architecture

See docs/ARCHITECTURE.md for detailed architecture documentation.

Key Design Decisions

• Provider-agnostic: Core never mentions specific providers (OpenAI, pgvector, etc.)
• Batch operations: Use embedBatch(), storeMany() for production performance
• Idempotent ingestion: Content hash check skips unchanged documents
• Separate database: Uses RAG_DB_* connection, not app's default
• Hard delete: No soft delete to avoid HNSW index bloat

Testing

# Run tests
composer test

# Run PHPStan
composer analyse

License

MIT License. See LICENSE.md.