{
"event": "PreToolUse",
"tool_name": "Read",
"tool_input": {
"file_path": "\/var\/www\/dev.campus.systemische-tools.de\/src\/Infrastructure\/AI\/ChatService.php",
"limit": 100
}
}
{
"tool_response": {
"type": "text",
"file": {
"filePath": "\/var\/www\/dev.campus.systemische-tools.de\/src\/Infrastructure\/AI\/ChatService.php",
"content": "<?php\n\ndeclare(strict_types=1);\n\nnamespace Infrastructure\\AI;\n\n\/\/ @responsibility: RAG-Pipeline für KI-Chat (Embedding, Suche, Antwort-Generierung)\n\nuse RuntimeException;\n\nfinal readonly class ChatService\n{\n \/**\n * Constructs a new ChatService instance.\n *\n * @param OllamaService $ollama Ollama service for embeddings and optional LLM\n * @param QdrantService $qdrant Qdrant service for vector search\n * @param ClaudeService $claude Claude service for high-quality LLM responses\n *\/\n public function __construct(\n private OllamaService $ollama,\n private QdrantService $qdrant,\n private ClaudeService $claude\n ) {\n }\n\n \/**\n * Executes a complete RAG chat pipeline.\n *\n * Performs the following steps:\n * 1. Generates an embedding vector for the question (if collections selected)\n * 2. Searches for similar documents in the vector database(s)\n * 3. Builds context from the most relevant chunks\n * 4. Generates an answer using the specified LLM model\n * 5. Extracts source information\n * 6. Assembles a structured response\n *\n * If no collections are selected, steps 1-3 and 5 are skipped (no RAG).\n *\n * @param string $question The user's question to answer\n * @param string $model The LLM model (claude-* or ollama:*)\n * @param array<string> $collections Qdrant collections to search (empty = no RAG)\n * @param int $limit Maximum number of document chunks to retrieve (default: 5)\n * @param string|null $stylePrompt Optional style prompt from author profile\n * @param string|null $customSystemPrompt Optional custom system prompt (replaces default if set)\n * @param float $temperature Sampling temperature 0.0-1.0 (default: 0.7)\n * @param int $maxTokens Maximum tokens in response (default: 4096)\n *\n * @return array{\n * question: string,\n * answer: string,\n * sources: array<int, array{title: string, score: float, content?: string}>,\n * model: string,\n * usage?: array{input_tokens: int, output_tokens: int},\n * chunks_used: int\n * } Complete chat response with answer, sources, and metadata\n *\n * @throws RuntimeException If embedding generation fails\n * @throws RuntimeException If vector search fails\n * @throws RuntimeException If LLM request fails\n *\n * @example\n * $chat = new ChatService($ollama, $qdrant, $claude);\n * \/\/ With RAG (multiple collections)\n * $result = $chat->chat('Was ist systemisches Coaching?', 'claude-opus-4-5-20251101', ['documents', 'mail'], 5);\n * \/\/ Without RAG (no collections)\n * $result = $chat->chat('Erkläre mir Python', 'claude-opus-4-5-20251101', [], 5);\n *\/\n public function chat(\n string $question,\n string $model = 'claude-opus-4-5-20251101',\n array $collections = [],\n int $limit = 5,\n ?string $stylePrompt = null,\n ?string $customSystemPrompt = null,\n float $temperature = 0.7,\n int $maxTokens = 4096\n ): array {\n $searchResults = [];\n $context = '';\n\n \/\/ Only perform RAG if collections are selected\n if ($collections !== []) {\n \/\/ Step 1: Generate embedding for the question\n try {\n $queryEmbedding = $this->ollama->getEmbedding($question);\n } catch (RuntimeException $e) {\n throw new RuntimeException(\n 'Embedding generation failed: ' . $e->getMessage(),\n 0,\n $e\n );\n }\n\n if ($queryEmbedding === []) {\n throw new RuntimeException('Embedding generation returned empty vector');\n }\n\n \/\/ Step 2: Search across all selected collections\n try {",
"numLines": 100,
"startLine": 1,
"totalLines": 337
}
}
}