omnicado/nette-ai-bundle

Nette integration for Symfony AI components

Maintainers

Package info

github.com/omnicado/nette-ai-bundle

pkg:composer/omnicado/nette-ai-bundle

Statistics

Installs: 0

Dependents: 0

Suggesters: 0

Stars: 0

Open Issues: 0

Security

Advisories: 0

Aikido package health analysis

v0.1.0 2026-03-04 12:57 UTC

Requires

Requires (Dev)

Suggests

MIT 307ea897c2b3af8257414ca3e2e06b9f6bd60d12

symfonynetteAgentaiopenaillmanthropic

This package is auto-updated.

Last update: 2026-03-04 12:59:21 UTC

README

Nette Framework integration for Symfony AI components.

Provides a seamless bridge between Symfony AI's framework-agnostic platform, agent, and tooling libraries and the Nette DI container, Tracy debugger, and console infrastructure.

Features

  • Platform registration — configure OpenAI, Anthropic, Google, Ollama (or custom) platforms via NEON
  • Agent wiring — define agents with model, system prompt, tools, and memory
  • Tool auto-discovery — services with #[AsTool] are automatically registered
  • Processor auto-discovery — services with #[AsInputProcessor] / #[AsOutputProcessor] are wired into agents
  • Security — restrict tool access with #[IsGrantedTool] based on Nette\Security\User roles
  • Memory — persist conversation memory via Nette\Caching\Cache
  • Tracy panel — inspect API calls, duration, token usage, and request/response payloads
  • Console commands — ai:platform:invoke and ai:agent:call

Requirements

  • PHP 8.2+
  • Nette 3.2+
  • Symfony AI Platform ^0.5

Installation

composer require omnicado/nette-ai-bundle

Install the platform bridge package for your provider:

# OpenAI
composer require symfony/ai-open-ai-platform

# Anthropic
composer require symfony/ai-anthropic-platform

# Google Gemini
composer require symfony/ai-google-platform

# Ollama (local)
composer require symfony/ai-ollama-platform

For agent support (tools, processors, memory):

composer require symfony/ai-agent

Configuration

Register the extension in your NEON config:

extensions:
    ai: Omnicado\NetteAiBundle\DI\AiExtension

Full configuration reference

ai:
    # Platform definitions (keyed by name)
    platforms:
        openai:
            api_key: %env.OPENAI_API_KEY%
            model: gpt-4o

        anthropic:
            api_key: %env.ANTHROPIC_API_KEY%
            model: claude-sonnet-4-5-20250929

        ollama:
            base_url: http://127.0.0.1:11434    # default
            model: llama3

        # Custom platform with explicit factory class
        custom:
            factory: App\AI\MyPlatformFactory
            api_key: %env.CUSTOM_API_KEY%
            model: my-model

    # Agent definitions (keyed by name)
    agents:
        default:
            platform: openai                      # required, references a platform name
            model: gpt-4o-mini                    # optional, overrides platform model
            system_prompt: "You are a helpful assistant."
            tools: auto                           # auto | [class list] | false
            memory: false                         # inject memory processor

        researcher:
            platform: anthropic
            system_prompt: "You are a research assistant."
            tools:
                - App\Tool\WebSearchTool
                - App\Tool\WikipediaTool
            memory: true

    # Memory provider (Nette Cache backed)
    memory:
        enabled: false
        cache_namespace: ai.memory

    # Security (role-based tool access)
    security:
        enabled: false

    # Tracy debug panel
    debug: %debugMode%

Configuration options

Option Type Default Description
platforms.<name>.factory ?string null Custom PlatformFactory class. Auto-resolved for openai, anthropic, google, ollama.
platforms.<name>.api_key ?string null API key (supports %env.*% parameters).
platforms.<name>.model ?string null Default model identifier.
platforms.<name>.base_url ?string null Base URL (ollama defaults to http://127.0.0.1:11434).
agents.<name>.platform string required References a platform name.
agents.<name>.model ?string null Model override. Falls back to the platform's model.
agents.<name>.system_prompt ?string null System prompt injected into conversations.
agents.<name>.tools 'auto'|string[]|false 'auto' Tool wiring mode (see Tools).
agents.<name>.memory bool false Enable memory injection for this agent.
memory.enabled bool false Register NetteCacheMemoryProvider.
memory.cache_namespace string 'ai.memory' Nette Cache namespace for memories.
security.enabled bool false Enable #[IsGrantedTool] role checks.
debug bool false Enable Tracy panel and event collection.

Autowiring

The first registered platform is autowired as PlatformInterface. The first registered agent is autowired as Agent. Additional platforms/agents are accessible by service name:

// Autowired (first registered)
public function __construct(
    private PlatformInterface $platform,
    private Agent $agent,
) {}

// By name
$container->getService('ai.platform.anthropic');
$container->getService('ai.agent.researcher');

Usage

Direct platform invocation

use Symfony\AI\Platform\Message\Message;
use Symfony\AI\Platform\Message\MessageBag;
use Symfony\AI\Platform\PlatformInterface;

final class MyService
{
    public function __construct(
        private PlatformInterface $platform,
    ) {}

    public function ask(string $question): string
    {
        $messages = new MessageBag(
            Message::forSystem('You are a helpful assistant.'),
            Message::ofUser($question),
        );

        $result = $this->platform->invoke('gpt-4o', $messages);

        return $result->getResult()->getContent();
    }
}

Using agents

use Symfony\AI\Agent\Agent;
use Symfony\AI\Platform\Message\Message;
use Symfony\AI\Platform\Message\MessageBag;

final class ChatService
{
    public function __construct(
        private Agent $agent,
    ) {}

    public function chat(string $userMessage): string
    {
        $messages = new MessageBag(
            Message::ofUser($userMessage),
        );

        $result = $this->agent->call($messages);

        return $result->getContent();
    }
}

Tools

Tools are PHP classes annotated with #[AsTool] from symfony/ai-agent. They are automatically discovered and injected into agents.

Creating a tool

use Symfony\AI\Agent\Toolbox\Attribute\AsTool;

#[AsTool('weather', 'Get current weather for a city')]
final class WeatherTool
{
    public function __invoke(string $city): string
    {
        // Fetch weather data...
        return "The weather in {$city} is sunny, 22°C.";
    }
}

Register the tool as a service:

services:
    - App\Tool\WeatherTool

Multi-method tools

A single class can expose multiple tools using the repeatable #[AsTool] attribute:

#[AsTool(name: 'weather_current', description: 'Get current weather', method: 'current')]
#[AsTool(name: 'weather_forecast', description: 'Get weather forecast', method: 'forecast')]
final class WeatherTool
{
    public function current(string $city): string { /* ... */ }
    public function forecast(string $city, int $days = 3): string { /* ... */ }
}

Tool wiring modes

agents:
    # Auto-discover all #[AsTool] services
    assistant:
        platform: openai
        tools: auto

    # Only specific tool classes
    researcher:
        platform: openai
        tools:
            - App\Tool\WebSearchTool
            - App\Tool\WikipediaTool

    # No tools
    simple:
        platform: openai
        tools: false

Security

Restrict tool access based on Nette\Security\User roles using the #[IsGrantedTool] attribute.

Setup

ai:
    security:
        enabled: true

Requires nette/security with a User service in the container.

Usage

use Omnicado\NetteAiBundle\Security\IsGrantedTool;
use Symfony\AI\Agent\Toolbox\Attribute\AsTool;

// Only users with the 'admin' role can use this tool
#[AsTool('delete_user', 'Deletes a user account')]
#[IsGrantedTool('admin')]
final class DeleteUserTool
{
    public function __invoke(int $userId): string
    {
        // ...
    }
}

// Users with either 'admin' or 'editor' role can use this tool
#[AsTool('publish_article', 'Publishes an article')]
#[IsGrantedTool('admin')]
#[IsGrantedTool('editor')]
final class PublishArticleTool
{
    public function __invoke(int $articleId): string
    {
        // ...
    }
}

// No #[IsGrantedTool] = available to everyone
#[AsTool('search', 'Searches documents')]
final class SearchTool
{
    public function __invoke(string $query): string
    {
        // ...
    }
}

When security is enabled:

  • getTools() filters out tools the current user cannot access (the LLM won't even know about them)
  • execute() returns an "Access denied" result if a tool is called without the required role

Memory

Persist conversation memories across requests using Nette Cache.

Setup

ai:
    memory:
        enabled: true
        cache_namespace: ai.memory    # optional

    agents:
        default:
            platform: openai
            memory: true              # enable for this agent

Programmatic usage

The NetteCacheMemoryProvider can also be used directly:

use Omnicado\NetteAiBundle\Memory\NetteCacheMemoryProvider;

final class MemoryService
{
    public function __construct(
        private NetteCacheMemoryProvider $memoryProvider,
    ) {}

    public function saveContext(string $fact): void
    {
        $this->memoryProvider->remember($fact);
    }

    public function clearMemory(): void
    {
        $this->memoryProvider->forget();
    }

    public function listMemories(): array
    {
        return $this->memoryProvider->all();
    }
}

When an agent has memory: true, a MemoryInputProcessor is prepended to its input processors. It loads stored memories and injects them into the system message before each call.

Processors

Custom input/output processors are automatically discovered and wired into agents.

Input processor

use Symfony\AI\Agent\Attribute\AsInputProcessor;
use Symfony\AI\Agent\Input;
use Symfony\AI\Agent\InputProcessorInterface;

// Applied to all agents
#[AsInputProcessor]
final class LoggingInputProcessor implements InputProcessorInterface
{
    public function processInput(Input $input): void
    {
        // Log, modify messages, inject context, etc.
    }
}

// Applied only to the 'researcher' agent, with high priority
#[AsInputProcessor(agent: 'researcher', priority: 100)]
final class ResearchContextProcessor implements InputProcessorInterface
{
    public function processInput(Input $input): void
    {
        // Add research-specific context
    }
}

Output processor

use Symfony\AI\Agent\Attribute\AsOutputProcessor;
use Symfony\AI\Agent\Output;
use Symfony\AI\Agent\OutputProcessorInterface;

#[AsOutputProcessor(agent: 'default', priority: 10)]
final class ResponseFilterProcessor implements OutputProcessorInterface
{
    public function processOutput(Output $output): void
    {
        // Filter, transform, or validate agent output
    }
}

Register processors as services:

services:
    - App\Processor\LoggingInputProcessor
    - App\Processor\ResearchContextProcessor
    - App\Processor\ResponseFilterProcessor

Processors are sorted by priority (higher = runs first). When agent is null, the processor applies to all agents.

Tracy Debug Panel

Enable the Tracy panel to monitor AI platform calls during development:

ai:
    debug: %debugMode%

The panel displays:

  • Bar tab — invocation count and total duration (e.g., "2 calls / 1234.5 ms")
  • Summary table — total API calls, total duration, prompt/completion/total token counts
  • Per-call details:
    • Model name
    • Duration in milliseconds
    • Token breakdown (prompt, completion, thinking, cached)
    • Input messages (role + content)
    • Output content (truncated to 2000 chars)
    • Options dump (temperature, etc.)

The panel hooks into Symfony AI Platform's event system via a lightweight PSR-14 dispatcher that intercepts InvocationEvent and ResultEvent.

Console Commands

Requires contributte/console or any other integration of symfony/console into Nette.

ai:platform:invoke

Invoke a platform directly from the command line:

# Basic usage
php bin/console ai:platform:invoke "What is PHP?"

# With options
php bin/console ai:platform:invoke "Explain quantum computing" \
    --model gpt-4o \
    --system "You are a physics professor." \
    --temperature 0.7
Option Short Description
prompt (argument) — The prompt to send (required)
--model -m Model to use (defaults to platform's model)
--system -s System prompt
--temperature -t Temperature (0.0–2.0)

ai:agent:call

Call an agent (with tools, memory, processors):

php bin/console ai:agent:call "Summarize the latest news about AI"

php bin/console ai:agent:call "Find weather in Prague" \
    --system "Always respond in Czech."
Option Short Description
prompt (argument) — The prompt to send (required)
--system -s System prompt

Architecture

src/
├── DI/
│   ├── AiExtension.php              # Main compiler extension
│   ├── AiEventDispatcher.php         # PSR-14 dispatcher for Tracy integration
│   └── Pass/
│       ├── ToolDiscoveryPass.php     # #[AsTool] discovery
│       └── ProcessorDiscoveryPass.php # #[AsInputProcessor] / #[AsOutputProcessor] discovery
├── Console/
│   ├── PlatformInvokeCommand.php     # ai:platform:invoke
│   └── AgentCallCommand.php          # ai:agent:call
├── Memory/
│   └── NetteCacheMemoryProvider.php  # MemoryProviderInterface over Nette Cache
├── Security/
│   ├── IsGrantedTool.php             # #[IsGrantedTool] attribute
│   └── NetteToolAccessChecker.php    # ToolboxInterface decorator with role checks
└── Tracy/
    ├── AiPanel.php                   # Tracy\IBarPanel implementation
    ├── AiPanelCollector.php          # Event data collector
    ├── InvocationRecord.php          # Single invocation value object
    └── templates/
        └── panel.phtml               # Panel HTML template

Minimal example

# config/common.neon

extensions:
    ai: Omnicado\NetteAiBundle\DI\AiExtension

ai:
    platforms:
        openai:
            api_key: %env.OPENAI_API_KEY%
            model: gpt-4o

    agents:
        default:
            platform: openai
            system_prompt: "You are a helpful assistant."
            tools: auto

    debug: %debugMode%

services:
    - App\Tool\WeatherTool
// app/Tool/WeatherTool.php

#[AsTool('weather', 'Get weather for a city')]
final class WeatherTool
{
    public function __invoke(string $city): string
    {
        return "Sunny, 22°C in {$city}.";
    }
}
// app/Presenters/ChatPresenter.php

final class ChatPresenter extends Nette\Application\UI\Presenter
{
    public function __construct(
        private Agent $agent,
    ) {}

    public function actionDefault(): void
    {
        $result = $this->agent->call(
            new MessageBag(Message::ofUser('What is the weather in Prague?')),
        );

        $this->sendJson(['response' => $result->getContent()]);
    }
}

License

MIT