Symfony AI in Action - SymfonyLive Berlin 2026

At first AI is a ton of buzzwords and hype, but how do we integrate AI in our existing products / software projects in a meaningful and realistic manner?
Let's have a look how Symfony AI enables us to bring AI-based or even agentic features, not only chatbots, into our PHP applications or scaled architecture.
If you are looking for input on how to approach AI, need ideas where to get started or find out what's realistic at first - let this talk be your inspiration!

This talk was given at SymfonyLive Berlin 2026 - https://live.symfony.com/2026-berlin/

Transcript

2. SymfonyLive Berlin 2025 Who's been around back then? A lot has happened - let's have a look! 2
3. Christopher Hertel Software Engineer in Berlin Focus on PHP, Symfony and GenAI Working on Tech Program of Active on Symfony AI & MCP PHP SDK 3
4. I N A C T I O N 4
5. Initiative 5
6. Initiative Converted LLM Chain Published July 2025 symfony/ai Symfony Standards Experimental 1.0 Release Soonish 6
7. Idea Enable AI Features Not only LLMs PHP Components Vendor Agnostic 7
8. Components Foundation Components Platform Agent Store Tools AI Bundle Framework Integration Third Party more than 60 Integrations already 60 MCP Bundle Framework Integration MCP SDK Server - Client 8
9. Platform 9
10. Platform Model Inference Abstraction composer require symfony/ai-platform 10
11. Platform Component Abstraction Layer for Model Platforms Central Interface for Inference Model Platform Bridges Multi-Modal & Streaming Support Structured Output Integrated Event System 11
12. Platform Component Contract Invokation DTOs Platform Provider Provider ... Model Inference Model Inference Contract ... Result DTOs Input - Text - Messages - Image - Audio Result - Text - Binary - Stream - Object Provider - 1:n Relation - Model Catalog - Capabilities - Models.dev Model Inference - Mostly HTTP - Remote & local - Coding Agents - "Native" PHP 12
13. Model Inference use Symfony\AI\Platform\Bridge\OpenAi\Factory; use Symfony\AI\Platform\Message\{Message, MessageBag}; $platform = Factory::createPlatform($apiKey); $input = new MessageBag( Message::forSystem('You are a pirate and you write funny.'), Message::ofUser('What is the Symfony framework?'), ); $result = $platform->invoke('gpt-5-mini', $input); echo $result->asText(); 5 6 7 8 9 10 11 12 13 14 15 16 17 13
14. OpenAI GPT use Symfony\AI\Platform\Bridge\OpenAi\Factory; $result = $platform->invoke('gpt-5-mini', $input); 5 use Symfony\AI\Platform\Message\{Message, MessageBag}; 6 7 $platform = Factory::createPlatform($apiKey); 8 9 $input = new MessageBag( 10 Message::forSystem('You are a pirate and you write funny.'), 11 Message::ofUser('What is the Symfony framework?'), 12 ); 13 14 15 16 echo $result->asText(); 17 14
15. Anthropic Claude use Symfony\AI\Platform\Bridge\Anthropic\Factory; $result = $platform->invoke('claude-sonnet-4-5-20250929', $input); 5 use Symfony\AI\Platform\Message\{Message, MessageBag}; 6 7 $platform = Factory::createPlatform($apiKey); 8 9 $input = new MessageBag( 10 Message::forSystem('You are a pirate and you write funny.'), 11 Message::ofUser('What is the Symfony framework?'), 12 ); 13 14 15 16 echo $result->asText(); 17 15
16. Google Gemini use Symfony\AI\Platform\Bridge\Gemini\Factory; $result = $platform->invoke('gemini-3-pro-preview', $input); 5 use Symfony\AI\Platform\Message\{Message, MessageBag}; 6 7 $platform = Factory::createPlatform($apiKey); 8 9 $input = new MessageBag( 10 Message::forSystem('You are a pirate and you write funny.'), 11 Message::ofUser('What is the Symfony framework?'), 12 ); 13 14 15 16 echo $result->asText(); 17 16
17. Mistral use Symfony\AI\Platform\Bridge\Mistral\Factory; $result = $platform->invoke('mistral-large-latest', $input); 5 use Symfony\AI\Platform\Message\{Message, MessageBag}; 6 7 $platform = Factory::createPlatform($apiKey); 8 9 $input = new MessageBag( 10 Message::forSystem('You are a pirate and you write funny.'), 11 Message::ofUser('What is the Symfony framework?'), 12 ); 13 14 15 16 echo $result->asText(); 17 17
18. Streaming $result = $platform->invoke('mistral-large-latest', $input, [ 'stream' => true, ]); foreach ($result->asTextStream() as $delta) { echo $delta; } $input = new MessageBag( 10 Message::forSystem('You are a pirate and you write funny.'), 11 Message::ofUser('What is the Symfony framework?'), 12 ); 13 14 15 16 17 18 19 20 21 18
19. Typed Streaming Deltas foreach ($result->asStream() as $delta) { if ($delta instanceof TextDelta) { echo $delta; } if ($delta instanceof ThinkingDelta) { echo "<info>" . $delta->getThinking() . "</info>"; } if ($delta instanceof ToolCallComplete) { $logger->info('tool invoked', ['name' => $delta->getName()]); } if ($delta instanceof TokenUsage) { $metrics->record($delta->getTotalTokens()); } } 19
20. Prompt Templates $input = new MessageBag( Message::forSystem('You are a helpful assistant.'), Message::ofUser(Template::string('Tell me about {topic}')), ); $result = $platform->invoke('gpt-5-mini', $input, [ 'template_vars' => ['topic' => 'Symfony AI'], ]); echo $result->asText(); 20
21. Audio, Image & PDF Input $input = new MessageBag( Message::ofUser( 'Describe the audio file.', Audio::fromDataUrl('data:audio/mpeg;base64,/9j/4AAQ...'), ), ); $result = $platform->invoke('gemini-2.5-flash', $input); echo $result->asText(); 10 11 12 13 14 15 16 17 18 19 Message::ofUser( 'Describe the audio file.', Audio::fromDataUrl('data:audio/mpeg;base64,/9j/4AAQ...'), ), $input = new MessageBag( 10 11 12 13 14 ); 15 16 $result = $platform->invoke('gemini-2.5-flash', $input); 17 18 echo $result->asText(); 19 21
22. Audio, Image & PDF Input Message::ofUser( 'What is visible on the image?', new ImageUrl('https://example.org/elephant.jpg'), ), $input = new MessageBag( 10 11 12 13 14 ); 15 16 $result = $platform->invoke('gemini-2.5-flash', $input); 17 18 echo $result->asText(); 19 22
23. Audio, Image & PDF Input Message::ofUser( 'Describe the content of the PDF file.', Document::fromFile('/path/to/document.pdf'), ), $input = new MessageBag( 10 11 12 13 14 ); 15 16 $result = $platform->invoke('gemini-2.5-flash', $input); 17 18 echo $result->asText(); 19 23
24. Binary Output $input = new MessageBag( Message::ofUser( 'Colorize the elephant in red.', Image::fromFile('/path/to/elephant.jpg'), ), ); $result = $platform->invoke('gemini-2.5-flash-image', $input); $result->asFile('/tmp/colorized.png'); → 24
25. Token Usage Also exposed as TokenUsage delta in streaming results. $result = $platform->invoke('gpt-5-mini', $input); $tokenUsage = $result->getMetadata()->get('token_usage'); echo $tokenUsage->getPromptTokens(); // 23 echo $tokenUsage->getCompletionTokens(); // 142 echo $tokenUsage->getTotalTokens(); // 165 25
26. Structured Output $input = new MessageBag( Message::forSystem('Help users as math tutor, step by step.'), Message::ofUser('how can I solve 8x + 7 = -23'), ); $result = $platform->invoke('mistral-small-latest', $input, [ 'response_format' => MathReasoning::class, ]); dump($result->asObject()); 26
27. Structured Output 27
28. Object Population $city = new City(name: 'Berlin'); $input = new MessageBag(Message::ofUser( Template::string('Provide missing information for {city.name}'), )); $platform->invoke('gpt-5-mini', $input, [ 'template_vars' => ['city' => $city], 'response_format' => $city, ]); echo $city->getMajor(); Kai Wegner 28
29. Embeddings Models Semantic Vectors are LLM's little sibling $result = $platform->invoke('text-embedding-3-small', $text); echo $result->asVectors(); // array of floats 29
30. Hugging Face Integration with Hugging Face's Inference Hub API use Symfony\AI\Platform\Bridge\HuggingFace\Task; $result = $platform->invoke('facebook/detr-resnet-50', $image, [ 'task' => Task::OBJECT_DETECTION, ]); echo $result->asObject(); // image classification result 30
31. Platform Bridges 35 Platform Bridges already, and more to come ... 31
32. Platform Events ModelRoutingEvent Before model resolution — observe or override provider routing InvocationEvent Before the platform call — mutate messages or options ResultEvent After the deferred result is created — react to the pending response StartEvent / DeltaEvent / CompleteEvent Streaming lifecycle for listeners on asStream() 32
33. AI Bundle Symfony Framework Integration composer require symfony/ai-bundle 33
34. AI Bundle Symfony Framework Integration composer require ai 34
35. Platform Configuration # config/packages/ai.yaml ai: platform: openai: api_key: '%env(OPENAI_API_KEY)%' anthropic: api_key: '%env(ANTHROPIC_API_KEY)%' # test command bin/console ai:platform:invoke openai gpt-5-mini "Hello Berlin" 35
36. Agent 36
37. Agent LLM ‐ Tool Call Iterator composer require symfony/ai-agent 37
38. Agent Component Multi-Step Interaction with Model and App Combining Platform & Model Collection of Messages/Prompts as Context Memory, Tools, and more Various Architectures From static workflows to autonomous orchestration 38
39. Agent Component Platform Model Inference Agent Tool Tool Tool Agent Agent defines - Instructions - Context - Tools - Model Model - Generates a response - Potentially invoking a tool Agent executes - Tool calls - Returns results to model Model - Generates a response - Ends with a stop response 39
40. Tool Calling use Doctrine\ORM\EntityManagerInterface; use Symfony\AI\Agent\Toolbox\Attribute\AsTool; #[AsTool('cookbook_create_recipe', 'Creates a new recipe in the cookbook.')] final readonly class CookbookCreateRecipe { public function __construct( private EntityManagerInterface $entityManager, ) {} /** * @param Recipe $recipe The recipe to create in the cookbook. */ public function __invoke(Recipe $recipe): string { // validate & persist recipe ... return 'Recipe created successfully!'; } } 40
41. Tool Argument Validation Requires symfony/validator to be installed. use Symfony\Component\Validator\Constraints as Assert; final readonly class Recipe { public function __construct( #[Assert\NotBlank] public string $name, #[Assert\GreaterThanOrEqual(1)] public int $servings, /** @var string[] */ #[Assert\All([new Assert\NotBlank()])] public array $ingredients, ) {} } 41
42. Agent Configuration # config/packages/ai.yaml ai: agent: wikipedia: model: name: 'gpt-5-mini' options: temperature: 0.2 prompt: 'You are a helpful research assistant.' tools: - 'Symfony\AI\Agent\Bridge\Wikipedia\Wikipedia' # test command bin/console ai:agent:call wikipedia Prompts can be a file, translation key, or service reference. 42
43. Agent Tool Bridges 13 Agent Tool Bridges already, and more to come ... 43
44. Agent Events ToolCallRequested Before tool runs — stoppable, listener can deny with a reason ToolCallArgumentsResolved Arguments deserialized — validator listener hooks in here ToolCallSucceeded / ToolCallFailed Per-tool result, with arguments, return value, or exception ToolCallsExecuted Aggregate after a batch of tool calls finishes 44
45. Human in the Loop With ToolCallRequested tool calls are stoppable! use Symfony\AI\Agent\Toolbox\Event\ToolCallRequested; $dispatcher->addListener(ToolCallRequested::class, function (ToolCall $toolCall = $event->getToolCall(); $approved = $io->confirm(sprintf( "Allow tool '%s' with args %s?", $toolCall->getName(), json_encode($toolCall->getArguments()), )); if (!$approved) { $event->deny('User denied tool execution.'); } }); 5 6 7 8 9 10 11 12 13 14 15 16 17 18 $dispatcher->addListener(ToolCallRequested::class, function (ToolCall use Symfony\AI\Agent\Toolbox\Event\ToolCallRequested; 5 6 7 $toolCall = $event->getToolCall(); 8 $approved = $io->confirm(sprintf( 9 "Allow tool '%s' with args %s?", 10 $toolCall->getName(), 11 json_encode($toolCall->getArguments()), 12 )); 13 14 if (!$approved) { 15 $event->deny('User denied tool execution.'); 16 } 17 }); 18 $approved = $io->confirm(sprintf( "Allow tool '%s' with args %s?", $toolCall->getName(), json_encode($toolCall->getArguments()), )); use Symfony\AI\Agent\Toolbox\Event\ToolCallRequested; 5 6 $dispatcher->addListener(ToolCallRequested::class, function (ToolCall 7 $toolCall = $event->getToolCall(); 8 9 10 11 12 13 14 if (!$approved) { 15 $event->deny('User denied tool execution.'); 16 } 17 }); 18 if (!$approved) { $event->deny('User denied tool execution.'); } use Symfony\AI\Agent\Toolbox\Event\ToolCallRequested; 5 6 $dispatcher->addListener(ToolCallRequested::class, function (ToolCall 7 $toolCall = $event->getToolCall(); 8 $approved = $io->confirm(sprintf( 9 "Allow tool '%s' with args %s?", 10 $toolCall->getName(), 11 json_encode($toolCall->getArguments()), 12 )); 13 14 15 16 17 }); 18 45
46. Multi-Agent Architectures Level of Autonomy & Context Sharing Implemented Workflows Solved in user land Subagents as Tools Defined subtask & isolated context Orchestration with Handoff Central orchestrator agent & handoffs with context to subagents 46
47. Subagent as Tool # config/packages/ai.yaml ai: # Platform Configuration ... agent: researcher: model: 'gpt-5' prompt: 'You do heavy research.' smalltalker: model: 'gpt-5-mini' tools: # Agent as tool 🤯 - agent: 'researcher' name: 'research_topic' description: 'Does research on a given topic.' 47
48. Orchestration # config/packages/ai.yaml ai: # Platform Configuration ... agent: orchestrator: { # agent config ... } technical: { # agent config ... } billing: { # agent config ... } fallback: { # agent config ... } multi_agent: support: orchestrator: 'orchestrator' handoffs: technical: ['bug', 'error', 'code', 'debug'] billing: ['invoice', 'payment', 'bill', 'order'] fallback: 'fallback' 48
49. Memory Register the provider as a service and reference it from ai.agent.*.memory.service. final class UserFactsMemory implements MemoryProviderInterface { public function __construct( private UserRepository $user, private Security $security, ) {} public function load(Input $input): array { $facts = $this->user->findFacts($this->security->getUser()); return $facts ? [ new Memory('## Known about the user'.\PHP_EOL.$facts) ] : []; } } 49
50. Memory Configuration # config/packages/ai.yaml ai: agent: blog: model: 'gpt-5-mini' prompt: 'You are a helpful booking agent.' tools: - 'App\Tool\BookingSearch' memory: service: 'App\Memory\UserFactsMemory' 50
51. Speech Agent Configuration Currently, a SpeechAgent decorator: STT → Agent → TTS # config/packages/ai.yaml ai: agent: voice: model: 'gpt-5-mini' prompt: | You are a friendly voice assistant. Keep answers concise. speech: speech_to_text_platform: 'ai.platform.openai' stt_model: 'whisper-1' text_to_speech_platform: 'ai.platform.elevenlabs' tts_model: 'eleven_multilingual_v2' tts_options: voice: 'Bill' 51
52. Store 52
53. Store Indexing- and Retrieval-Pipeline composer require symfony/ai-store 53
54. Store Component Vector Store Abstraction Layer Lifecycle Management Adding & Querying for Documents Document Indexing Pipeline 1. Loading 2. Filtering 3. Transforming 4. Vectorizing 54
55. Store Configuration # config/packages/ai.yaml ai: store: postgres: blog: dbal_connection: 'doctrine.dbal.default_connection' table_name: 'blog_vectors' vectorizer: openai: platform: 'ai.platform.openai' model: 'text-embedding-3-small' retriever: blog: store: 'ai.store.postgres.blog' vectorizer: 'ai.vectorizer.openai' 55
56. Store Bridges 24 Store Bridges already, more to come, some to die ... 56
57. Indexing Pipeline $vectorizer = new Vectorizer($platform, 'text-embedding-3-small'); $indexer = new SourceIndexer( loader: new TextFileLoader(), processor: new DocumentProcessor( vectorizer: $vectorizer, store: $store, transformers: [ new TextReplaceTransformer(search: '## Plot', replace: '## Synopsis'), new TextSplitTransformer(chunkSize: 500, overlap: 100), ], ), ); $indexer->index(['/docs/gladiator.md', '/docs/inception.md']); 57
58. Indexer Configuration # config/packages/ai.yaml ai: indexer: blog: loader: 'Symfony\AI\Store\Document\Loader\RssFeedLoader' source: 'https://feeds.feedburner.com/symfony/blog' transformers: - 'Symfony\AI\Store\Document\Transformer\TextSplitTransformer' vectorizer: 'ai.vectorizer.openai' store: 'ai.store.postgres.blog' # indexing command bin/console ai:store:index blog 58
59. Loaders & Transformers Loaders TextFileLoader MarkdownLoader CsvLoader JsonFileLoader RssFeedLoader RstLoader / RstToctreeLoader InMemoryLoader Transformers TextSplitTransformer TextReplaceTransformer TextTrimTransformer ChunkDelayTransformer SummaryGeneratorTransformer ChainTransformer 59
60. Retriever & Queries $retriever = new Retriever($store, $vectorizer); // High-level: text in, documents out (wraps TextQuery) $documents = $retriever->retrieve('mafia movie'); // Low-level: pick the Query shape yourself $store->query(new TextQuery('mafia movie')); $store->query(new VectorQuery($vectorizer->vectorize('mafia movie'))); $store->query(new HybridQuery( vector: $vectorizer->vectorize('mafia movie'), text: 'mafia movie', semanticRatio: 0.7, // 70% semantic, 30% keyword )); 5 6 7 8 9 10 11 12 13 14 15 16 17 $retriever = new Retriever($store, $vectorizer); 5 6 // High-level: text in, documents out (wraps TextQuery) 7 $documents = $retriever->retrieve('mafia movie'); 8 9 // Low-level: pick the Query shape yourself 10 $store->query(new TextQuery('mafia movie')); 11 $store->query(new VectorQuery($vectorizer->vectorize('mafia movie'))); 12 $store->query(new HybridQuery( 13 vector: $vectorizer->vectorize('mafia movie'), 14 text: 'mafia movie', 15 semanticRatio: 0.7, // 70% semantic, 30% keyword 16 )); 17 $documents = $retriever->retrieve('mafia movie'); $retriever = new Retriever($store, $vectorizer); 5 6 // High-level: text in, documents out (wraps TextQuery) 7 8 9 // Low-level: pick the Query shape yourself 10 $store->query(new TextQuery('mafia movie')); 11 $store->query(new VectorQuery($vectorizer->vectorize('mafia movie'))); 12 $store->query(new HybridQuery( 13 vector: $vectorizer->vectorize('mafia movie'), 14 text: 'mafia movie', 15 semanticRatio: 0.7, // 70% semantic, 30% keyword 16 )); 17 $store->query(new TextQuery('mafia movie')); $retriever = new Retriever($store, $vectorizer); 5 6 // High-level: text in, documents out (wraps TextQuery) 7 $documents = $retriever->retrieve('mafia movie'); 8 9 // Low-level: pick the Query shape yourself 10 11 $store->query(new VectorQuery($vectorizer->vectorize('mafia movie'))); 12 $store->query(new HybridQuery( 13 vector: $vectorizer->vectorize('mafia movie'), 14 text: 'mafia movie', 15 semanticRatio: 0.7, // 70% semantic, 30% keyword 16 )); 17 $store->query(new VectorQuery($vectorizer->vectorize('mafia movie'))); $retriever = new Retriever($store, $vectorizer); 5 6 // High-level: text in, documents out (wraps TextQuery) 7 $documents = $retriever->retrieve('mafia movie'); 8 9 // Low-level: pick the Query shape yourself 10 $store->query(new TextQuery('mafia movie')); 11 12 $store->query(new HybridQuery( 13 vector: $vectorizer->vectorize('mafia movie'), 14 text: 'mafia movie', 15 semanticRatio: 0.7, // 70% semantic, 30% keyword 16 )); 17 $store->query(new HybridQuery( vector: $vectorizer->vectorize('mafia movie'), text: 'mafia movie', semanticRatio: 0.7, // 70% semantic, 30% keyword )); $retriever = new Retriever($store, $vectorizer); 5 6 // High-level: text in, documents out (wraps TextQuery) 7 $documents = $retriever->retrieve('mafia movie'); 8 9 // Low-level: pick the Query shape yourself 10 $store->query(new TextQuery('mafia movie')); 11 $store->query(new VectorQuery($vectorizer->vectorize('mafia movie'))); 12 13 14 15 16 17 60
61. Store Events PreQueryEvent Before documents are retrieved — enhance or rewrite the query PostQueryEvent After documents are retrieved — filter, rerank, or enrich results 61
62. SimilaritySearch Configuration # config/packages/ai.yaml ai: agent: blog: model: 'gpt-5-mini' prompt: 'Answer using the similarity_search tool.' tools: - 'Symfony\AI\Agent\Bridge\SimilaritySearch\SimilaritySearch' 62
63. Retrieval Augmented Generation 63
64. Demo 64
65. Symfony AI Demo Application github.com/symfony/ai-demo 65
66. Roadmap 66
67. Roadmap Agent API Changes Release v1.0 Deep MCP Support Doctrine Support++ 67
68. Symfony AI Roadmap Multi-Agent Architectures Agent Traces Evaluations & Benchmarking Parallelization / Orchestration Support for Protocols like A2A & AG-UI Tool Discovery Built-in Mercure Support Better Docs, Cookbook, and Demo ... 68
69. Only my plans, but what about you? 69
70. Already over 120 contributors! Thank You All! 70
71. Get Started with Symfony AI Checkout symfony/ai repository Especially examples/ and demo/ folders Build your first feature & give feedback See you on GitHub.com! 71
72. THANK YOU • QUESTI ONS? 72