Agentic Applications with Symfony Christopher Hertel ‐ SymfonyLive Berlin 2025 1 

Agenda Theory & Terminology Models, Context, RAG, and more Evolution from LLM to Agentic Code Examples with Symfony Model Context Protocol Distributed Agentic Applications 2 

Christopher Hertel Software Engineer Focus on PHP, Symfony and GenAI Working on Tech Program of 3 

Slides Scan and follow the slides online 4 

Theory & Terminology Models, Context, RAG, and more 5 

Models Language Model Trained to generate text based on text input Strong Competition & Fast Innovation OpenAI, Google, Meta, Anthropic, Mistral, and more Various Capabilities Streaming, structured output, multi-modal, real-time, tool calling Reasoning Model Decision making, less prompt engineering & more autonomous 6 

Platforms Model Vendor OpenAI, Anthropic, Mistral, Google, ... General Cloud Provider Azure (GitHub), Google, AWS, ... Specialized Cloud Provider HuggingFace, Replicate, OpenRouter, ... Local Runtimes GPT4All, Ollama, ONNX Runtime, ... 7 

Chain Chain of Calls Designed for multi-step interaction with models Combination of Things Platform, Models, System Prompts, Capabilities, etc. LangChain Most Famous LLM framework (Pyhton & TypeScript) 8 

Plain Model Chain9 

Context Stateless API Pre-trained, fine-tuned, but API "forgets" Limitation Size of context varies per model, but limited Powerful Control Mechanism Efficient to steer and evaluate the models generation 10 

Context Example 11 

RAG Architecture Retrieval Augmented Generation Dynamically inject information into context Improves Accuracy and Relevance More context for the model to work with, based on input Vector Stores Classic RAG combines embedding vectors and similarity search Improvement Techniques Pre-processing, Multi-Query, Reranking, and more 12 

Simple RAG13 

Tool Calling App Provides Tools Exposed to tool via JSON Schema definition as part of API call Model Invokes Tool As a intermediate response the model invokes a tool (or multiple) Tool Result Result of tool call provided as next call to model Instrumentation Log and analyze tool calls, model invocations, and results 14 

Tool Calling15 

Agent vs. Agentic 16 

PHP Libraries Listicle on GitHub LLPhant theodo-group/llphant Prism prism-php/prism LLM Chain php-llm/llm-chain Today 17 

Evolution from LLM to Agent Code Examples with PHP & Symfony 18 

Install LLM Chain composer require php-llm/llm-chain 19 

Plain Model Chain20 

Plain Model Chain use PhpLlm\LlmChain\Bridge\OpenAI\{GPT, PlatformFactory}; use PhpLlm\LlmChain\Chain; use PhpLlm\LlmChain\Model\Message\{Message, MessageBag}; $platform = PlatformFactory::create($apiKey); $chain = new Chain($platform, new GPT()); $messages = new MessageBag( Message::forSystem('You are a pirate and you write funny.'), Message::ofUser('What is the Symfony framework?'), ); $response = $chain->call($messages); echo $response->getContent(); 21 

Symfony Integration composer require php-llm/llm-chain-bundle 22 

Bundle Configuration # config/packages/llm_chain.yaml llm_chain: platform: openai: # or anthropic, azure, google, more to come ... api_key: '%env(OPENAI_API_KEY)%' chain: default: model: name: 'GPT' version: 'gpt-4o-mini' system_prompt: 'You are a pirate and you write funny.' 23 

Usage in Symfony final readonly class SomeService { public function __construct(private ChainInterface $chain) { } public function someMethod(string $message): string { $messages = new MessageBag(Message::ofUser($message)); $response = $this->chain->call($messages); return $response->getContent(); } } 24 

Simple RAG25 

Tool Calling26 

Tool Calling use PhpLlm\LlmChain\Chain\Toolbox\Attribute\AsTool; use Symfony\Component\Clock\ClockInterface; #[AsTool('clock', description: 'Returns the current date and time.')] final readonly class Clock { public function __construct(private ClockInterface $clock) { } public function __invoke(): string { return $this->clock->now()->format('Y-m-d H:i:s'); } } That's it already! 27 

Agentic RAG28 

Bundle Configuration # config/packages/llm_chain.yaml llm_chain: # Platform & Chain Configuration ... store: chroma_db: # or Azure Search, MongoDB, Pinecone, more to come ... symfonyblog: { collection: 'symfony_blog' } embedder: default: model: { name: 'Embeddings', version: 'text-embedding-ada-002' } services: _defaults: { autowire: true, autoconfigure: true } PhpLlm\LlmChain\Chain\Toolbox\Tool\SimilaritySearch: ~ 29 

Agentic Application30 

Bundle Configuration # config/packages/llm_chain.yaml llm_chain: # Platform Configuration ... chain: wikipedia: model: name: 'GPT' version: 'gpt-4o-mini' system_prompt: | Please answer the users question based on information of Wikipedia and provide a link to the article. tools: - 'PhpLlm\LlmChain\Chain\Toolbox\Tool\Wikipedia' 31 

AI Tools32 

Bundle Configuration Stack and isolate chains to build a more complex chain. # config/packages/llm_chain.yaml llm_chain: # Platform Configuration ... chain: complex_chain: # Model Configuration ... tools: # Chain in chain 🤯 - service: 'llm_chain.chain.blog' name: 'symfony_blog' description: 'Can answer questions based on the Symfony blog.' is_chain: true 33 

Model Context Protocol Distributed Agentic Applications 34 

Model Context Protocol (MCP) Open Sourced Protocol by Anthropic Only release late 2024, already exploded to everywhere Standardizes Five Aspects Resources, Prompts, Tools, Sampling, Roots Two Main Components Defines MCP Server and MCP Client JSON-RPC with Two Transports STDIO and Server-Sent-Events 35 

General Architecture36 

Agentic Application37 

Model Context Protocol38 

PHP MCP SDK No official SDK for PHP 39 

But ... composer require php-llm/mcp-bundle 40 

Bundle Configuration # config/packages/mcp.yaml mcp: app: 'app' # Application name to be exposed to clients version: '1.0.0' # Application version to be exposed to clients # Configure this application to act as a MCP server # For now ONLY TOOLS client_transports: stdio: true # Enable STDIO via command sse: true # Enable Server-Sent Event via controller # Configure MCP servers to be used by this application # NOT IMPLEMENTED YET servers: name: transport: 'stdio' # Transport method to use, either 'stdio' or 'sse' stdio: command: 'php /path/bin/console mcp' # Command to execute to start arguments: [] # Arguments to pass to the command sse: url: 'http://localhost:8000/sse' # URL to SSE endpoint of MCP serv 41 

Work In Progress But let's have a look ... 42 

Thank You! Any Questions? 43 

Useful Links LLM Chain: github.com/php-llm/llm-chain Bundle: github.com/php-llm/llm-chain-bundle MCP SDK (WIP): github.com/php-llm/mcp-sdk Bundle (WIP): github.com/php-llm/mcp-bundle Other Libs: github.com/php-llm/ecosystem Model Context Protocol: modelcontextprotocol.io 44