Mindwave: Production AI Utilities for Laravel

The working developer's AI toolkit - Long prompts, streaming, tracing, and context discovery made simple.

v1.0.0 Released - All 4 pillars complete with 1300+ tests.

Experimental - This package is under active development. APIs may change. Use in production at your own risk.

What is Mindwave?

Mindwave is a Laravel package that provides production-grade AI utilities for building LLM-powered features. Unlike complex agent frameworks, Mindwave focuses on practical tools that Laravel developers actually need:

• ✅ Auto-fit long prompts to any model's context window
• ✅ Stream LLM responses with 3 lines of code (SSE/EventSource)
• ✅ OpenTelemetry tracing with database storage for costs, tokens, and performance
• ✅ Ad-hoc context discovery from your database/CSV using TNTSearch

Why Mindwave?

Not another agent framework. Just batteries-included utilities for shipping AI features fast.

// Write long prompts, Mindwave auto-fits to model limits
Mindwave::prompt()
    ->section('system', $instructions)
    ->section('context', $largeDocument, priority: 50, shrinker: 'summarize')
    ->section('user', $question)
    ->fit()  // Auto-trims to context window
    ->run();

// Stream responses in 3 lines (backend)
return Mindwave::stream($prompt)->respond();

// View traces and costs
$traces = MindwaveTrace::expensive(0.10)->with('spans')->get();

// Pull context from your DB on-the-fly
Mindwave::prompt()
    ->context(TntSearchSource::fromEloquent(User::query(), fn($u) => "Name: {$u->name}"))
    ->ask('Who has Laravel expertise?');

Installation

Install via Composer:

composer require mindwave/mindwave

Publish the config files:

php artisan vendor:publish --tag="mindwave-config"

Run migrations for tracing (optional but recommended):

php artisan migrate

Quick Start

1. Basic LLM Chat

use Mindwave\Mindwave\Facades\Mindwave;

$response = Mindwave::llm()->chat([
    ['role' => 'system', 'content' => 'You are a helpful assistant.'],
    ['role' => 'user', 'content' => 'Explain Laravel in one sentence.'],
]);

echo $response->content;

2. Streaming Responses

Backend:

use Mindwave\Mindwave\Facades\Mindwave;

Route::get('/chat', function (Request $request) {
    return Mindwave::stream($request->input('message'))
        ->model('gpt-4')
        ->respond();
});

Frontend:

const stream = new EventSource('/chat?message=' + encodeURIComponent(question));
stream.onmessage = e => output.textContent += e.data;
stream.addEventListener('done', () => stream.close());

3. Auto-Fit Long Prompts

use Mindwave\Mindwave\Facades\Mindwave;

// Automatically handles token limits
Mindwave::prompt()
    ->reserveOutputTokens(500)
    ->section('system', 'You are an expert analyst', priority: 100)
    ->section('documentation', $longDocContent, priority: 50, shrinker: 'summarize')
    ->section('history', $conversationHistory, priority: 75)
    ->section('user', $userQuestion, priority: 100)
    ->fit()  // Trims to model's context window
    ->run();

4. View Costs & Traces

use Mindwave\Mindwave\Observability\Models\Trace;
use Mindwave\Mindwave\Observability\Models\Span;

// Find expensive traces
$expensive = Trace::where('estimated_cost', '>', 0.10)
    ->with('spans')
    ->orderByDesc('created_at')
    ->get();

// Find slow LLM calls
$slow = Span::where('operation_name', 'chat')
    ->where('duration', '>', 5_000_000_000) // 5 seconds in nanoseconds
    ->with('trace')
    ->get();

// Daily cost summary
$dailyCosts = Trace::selectRaw('
        DATE(created_at) as date,
        COUNT(*) as total_traces,
        SUM(estimated_cost) as total_cost,
        SUM(total_input_tokens) as input_tokens,
        SUM(total_output_tokens) as output_tokens
    ')
    ->groupBy('date')
    ->orderByDesc('date')
    ->get();

5. Ad-Hoc Context Discovery

use Mindwave\Mindwave\Context\Sources\TntSearchSource;

// Search your database on-the-fly
Mindwave::prompt()
    ->context(
        TntSearchSource::fromEloquent(
            Product::where('active', true),
            fn($p) => "Product: {$p->name}, Price: {$p->price}"
        )
    )
    ->ask('What products under $50 do you have?');

// Or from CSV files
Mindwave::prompt()
    ->context(TntSearchSource::fromCsv('data/knowledge-base.csv'))
    ->ask('How do I reset my password?');

Core Features

🧩 Prompt Composer

Automatically manage context windows with priority-based section trimming:

• Token budgeting - Reserve tokens for output, auto-fit sections
• Smart shrinkers - Summarize, truncate, or compress content
• Priority system - Keep important sections, trim less critical ones
• Multi-model support - Works with GPT-4, Claude, Mistral, etc.

🌊 Streaming (SSE)

Production-ready Server-Sent Events streaming:

• 3-line setup - Backend and frontend
• Proper headers - Works with Nginx/Apache out of the box
• Connection monitoring - Handles client disconnects
• Error handling - Graceful failure and retry

📊 OpenTelemetry Tracing

Industry-standard observability with GenAI semantic conventions:

• Automatic tracing - All LLM calls tracked (zero configuration)
• Database storage - Query traces via Eloquent models
• OTLP export - Send to Jaeger, Grafana, Datadog, Honeycomb, etc.
• Cost tracking - Automatic cost estimation per call
• Token usage - Input/output/total tokens tracked
• PII protection - Configurable message capture and redaction
• Artisan commands - Export, prune, and analyze traces

Quick Start:

// 1. Enable tracing in .env
// MINDWAVE_TRACING_ENABLED=true

// 2. LLM calls are automatically traced
$response = Mindwave::llm()->generateText('Hello!');

// 3. Query traces
use Mindwave\Mindwave\Observability\Models\Trace;

$expensive = Span::where('cost_usd', '>', 0.10)
    ->orderBy('cost_usd', 'desc')
    ->get();

📖 Complete Tracing Guide - Querying, cost analysis, custom spans, OTLP setup

📐 Architecture Documentation - Technical deep dive

🔍 TNTSearch Context Discovery

Pull context from your application data without complex RAG setup:

• No infrastructure - Pure PHP, no external services
• Multiple sources - Eloquent, arrays, CSV files, VectorStores
• Fast indexing - Ephemeral indexes with automatic cleanup
• BM25 ranking - Industry-standard relevance scoring
• Auto-query extraction - Automatically extracts search terms from user messages
• OpenTelemetry tracing - Track search performance and results

Quick Example:

use Mindwave\Mindwave\Context\Sources\TntSearch\TntSearchSource;
use Mindwave\Mindwave\Context\ContextPipeline;

// Search Eloquent models
$userSource = TntSearchSource::fromEloquent(
    User::where('active', true),
    fn($u) => "Name: {$u->name}, Skills: {$u->skills}"
);

// Search CSV files
$docsSource = TntSearchSource::fromCsv('data/knowledge-base.csv');

// Combine multiple sources
$pipeline = (new ContextPipeline)
    ->addSource($userSource)
    ->addSource($docsSource)
    ->deduplicate()  // Remove duplicates
    ->rerank();      // Sort by relevance

// Use in prompt (query auto-extracted from user message)
Mindwave::prompt()
    ->context($pipeline, limit: 5)
    ->section('user', 'Who has Laravel expertise?')
    ->run();

📖 Complete Context Discovery Guide - All source types, pipelines, advanced features

Configuration

LLM Configuration

// config/mindwave-llm.php
return [
    'default' => env('MINDWAVE_LLM_DRIVER', 'openai'),
    
    'llms' => [
        'openai' => [
            'api_key' => env('OPENAI_API_KEY'),
            'model' => env('OPENAI_MODEL', 'gpt-4-turbo'),
            'max_tokens' => 4096,
            'temperature' => 0.7,
        ],
        'mistral' => [
            'api_key' => env('MISTRAL_API_KEY'),
            'model' => env('MISTRAL_MODEL', 'mistral-large-latest'),
        ],
    ],
];

Tracing Configuration

// config/mindwave-tracing.php
return [
    'enabled' => env('MINDWAVE_TRACING_ENABLED', true),
    
    'database' => [
        'enabled' => true,  // Store in database
    ],
    
    'otlp' => [
        'enabled' => env('MINDWAVE_TRACE_OTLP_ENABLED', false),
        'endpoint' => env('OTEL_EXPORTER_OTLP_ENDPOINT', 'http://localhost:4318'),
    ],
    
    'capture_messages' => false,  // PII protection
    'retention_days' => 30,
];

Artisan Commands

# Export traces to CSV/JSON
php artisan mindwave:export-traces --since=yesterday --format=csv

# Prune old traces
php artisan mindwave:prune-traces --older-than=30days

# View trace statistics
php artisan mindwave:trace-stats

# View TNTSearch index statistics
php artisan mindwave:index-stats

# Clear old TNTSearch indexes (default: 24 hours)
php artisan mindwave:clear-indexes