{
"event": "PreToolUse",
"tool_name": "Edit",
"tool_input": {
"file_path": "\/var\/www\/dev.campus.systemische-tools.de\/src\/Infrastructure\/AI\/ChatService.php",
"old_string": " \/**\n * Builds a context string from search results.\n *\n * Concatenates the content from multiple search results into a single\n * context string, respecting a maximum character limit. Each chunk is\n * labeled with its source document title.\n *\n * @param array<int, array{id: int|string, score: float, payload: array<string, mixed>}> $searchResults Vector search results\n * @param int $maxTokens Maximum tokens to include (default: 3000)\n *\n * @return string The built context string\n *\/\n private function buildContext(array $searchResults, int $maxTokens = 3000): string\n {\n $contextParts = [];\n $totalChars = 0;\n $maxChars = $maxTokens * 4; \/\/ Approximate: 1 token ~ 4 characters\n\n foreach ($searchResults as $index => $result) {\n $payload = $result['payload'];\n \/\/ Support both payload schemas: documents + dokumentation_chunks\n $content = (string) ($payload['content'] ?? $payload['content_preview'] ?? '');\n $docTitle = (string) ($payload['document_title'] ?? $payload['title'] ?? 'Unbekannt');\n\n \/\/ Check if adding this chunk would exceed the limit\n if ($totalChars + strlen($content) > $maxChars) {\n break;\n }\n\n $contextParts[] = sprintf('[Quelle %d: %s]%s%s', $index + 1, $docTitle, \"\\n\", $content);\n $totalChars += strlen($content);\n }\n\n return implode(\"\\n\\n---\\n\\n\", $contextParts);\n }",
"new_string": " \/**\n * Builds a context string from search results.\n *\n * Concatenates the content from multiple search results into a single\n * context string, respecting a maximum character limit. Each chunk is\n * labeled with its source document title.\n *\n * Optionally prepends a semantic summary with extracted entities and taxonomy.\n *\n * @param array<int, array{id: int|string, score: float, payload: array<string, mixed>}> $searchResults Vector search results\n * @param int $maxTokens Maximum tokens to include (default: 3000)\n * @param string $semanticSummary Optional semantic summary to prepend\n *\n * @return string The built context string\n *\/\n private function buildContext(array $searchResults, int $maxTokens = 3000, string $semanticSummary = ''): string\n {\n $contextParts = [];\n $totalChars = 0;\n $maxChars = $maxTokens * 4; \/\/ Approximate: 1 token ~ 4 characters\n\n \/\/ Prepend semantic summary if available\n if ($semanticSummary !== '') {\n $contextParts[] = \"[Semantischer Kontext]\\n\" . $semanticSummary;\n $totalChars += strlen($semanticSummary);\n }\n\n foreach ($searchResults as $index => $result) {\n $payload = $result['payload'];\n \/\/ Support both payload schemas: documents + dokumentation_chunks\n $content = (string) ($payload['content'] ?? $payload['content_preview'] ?? '');\n $docTitle = (string) ($payload['document_title'] ?? $payload['title'] ?? 'Unbekannt');\n\n \/\/ Check if adding this chunk would exceed the limit\n if ($totalChars + strlen($content) > $maxChars) {\n break;\n }\n\n \/\/ Build chunk header with optional entity info\n $entities = $payload['entities'] ?? [];\n $entityInfo = '';\n if ($entities !== []) {\n $entityNames = array_slice(array_column($entities, 'name'), 0, 3);\n $entityInfo = ' | Enthält: ' . implode(', ', $entityNames);\n }\n\n $contextParts[] = sprintf('[Quelle %d: %s%s]%s%s', $index + 1, $docTitle, $entityInfo, \"\\n\", $content);\n $totalChars += strlen($content);\n }\n\n return implode(\"\\n\\n---\\n\\n\", $contextParts);\n }"
}
}
{
"tool_response": {
"filePath": "\/var\/www\/dev.campus.systemische-tools.de\/src\/Infrastructure\/AI\/ChatService.php",
"oldString": " \/**\n * Builds a context string from search results.\n *\n * Concatenates the content from multiple search results into a single\n * context string, respecting a maximum character limit. Each chunk is\n * labeled with its source document title.\n *\n * @param array<int, array{id: int|string, score: float, payload: array<string, mixed>}> $searchResults Vector search results\n * @param int $maxTokens Maximum tokens to include (default: 3000)\n *\n * @return string The built context string\n *\/\n private function buildContext(array $searchResults, int $maxTokens = 3000): string\n {\n $contextParts = [];\n $totalChars = 0;\n $maxChars = $maxTokens * 4; \/\/ Approximate: 1 token ~ 4 characters\n\n foreach ($searchResults as $index => $result) {\n $payload = $result['payload'];\n \/\/ Support both payload schemas: documents + dokumentation_chunks\n $content = (string) ($payload['content'] ?? $payload['content_preview'] ?? '');\n $docTitle = (string) ($payload['document_title'] ?? $payload['title'] ?? 'Unbekannt');\n\n \/\/ Check if adding this chunk would exceed the limit\n if ($totalChars + strlen($content) > $maxChars) {\n break;\n }\n\n $contextParts[] = sprintf('[Quelle %d: %s]%s%s', $index + 1, $docTitle, \"\\n\", $content);\n $totalChars += strlen($content);\n }\n\n return implode(\"\\n\\n---\\n\\n\", $contextParts);\n }",
"newString": " \/**\n * Builds a context string from search results.\n *\n * Concatenates the content from multiple search results into a single\n * context string, respecting a maximum character limit. Each chunk is\n * labeled with its source document title.\n *\n * Optionally prepends a semantic summary with extracted entities and taxonomy.\n *\n * @param array<int, array{id: int|string, score: float, payload: array<string, mixed>}> $searchResults Vector search results\n * @param int $maxTokens Maximum tokens to include (default: 3000)\n * @param string $semanticSummary Optional semantic summary to prepend\n *\n * @return string The built context string\n *\/\n private function buildContext(array $searchResults, int $maxTokens = 3000, string $semanticSummary = ''): string\n {\n $contextParts = [];\n $totalChars = 0;\n $maxChars = $maxTokens * 4; \/\/ Approximate: 1 token ~ 4 characters\n\n \/\/ Prepend semantic summary if available\n if ($semanticSummary !== '') {\n $contextParts[] = \"[Semantischer Kontext]\\n\" . $semanticSummary;\n $totalChars += strlen($semanticSummary);\n }\n\n foreach ($searchResults as $index => $result) {\n $payload = $result['payload'];\n \/\/ Support both payload schemas: documents + dokumentation_chunks\n $content = (string) ($payload['content'] ?? $payload['content_preview'] ?? '');\n $docTitle = (string) ($payload['document_title'] ?? $payload['title'] ?? 'Unbekannt');\n\n \/\/ Check if adding this chunk would exceed the limit\n if ($totalChars + strlen($content) > $maxChars) {\n break;\n }\n\n \/\/ Build chunk header with optional entity info\n $entities = $payload['entities'] ?? [];\n $entityInfo = '';\n if ($entities !== []) {\n $entityNames = array_slice(array_column($entities, 'name'), 0, 3);\n $entityInfo = ' | Enthält: ' . implode(', ', $entityNames);\n }\n\n $contextParts[] = sprintf('[Quelle %d: %s%s]%s%s', $index + 1, $docTitle, $entityInfo, \"\\n\", $content);\n $totalChars += strlen($content);\n }\n\n return implode(\"\\n\\n---\\n\\n\", $contextParts);\n }",
"originalFile": "<?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 * @param ScoringService $scoring Scoring service for weighted result ranking\n * @param SemanticEnrichmentService|null $semantic Optional semantic enrichment (graceful degradation)\n *\/\n public function __construct(\n private OllamaService $ollama,\n private QdrantService $qdrant,\n private ClaudeService $claude,\n private ScoringService $scoring,\n private ?SemanticEnrichmentService $semantic = null\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 {\n $searchResults = $this->searchMultipleCollections($queryEmbedding, $collections, $limit);\n } catch (RuntimeException $e) {\n throw new RuntimeException(\n 'Vector search failed: ' . $e->getMessage(),\n 0,\n $e\n );\n }\n\n \/\/ Step 2b: Semantic enrichment (graceful degradation - works without)\n $semanticSummary = '';\n if ($searchResults !== [] && $this->semantic !== null) {\n $searchResults = $this->semantic->enrichSearchResults($searchResults);\n $semanticSummary = $this->semantic->buildSemanticSummary($searchResults);\n }\n\n \/\/ Step 3: Build context from search results (if any found)\n if ($searchResults !== []) {\n $context = $this->buildContext($searchResults, 3000, $semanticSummary);\n }\n }\n\n \/\/ Step 4: Parse model string and generate answer\n $isOllama = str_starts_with($model, 'ollama:');\n $isClaude = str_starts_with($model, 'claude-');\n $hasContext = $context !== '';\n\n if ($isClaude) {\n try {\n \/\/ Build prompt: RAG with context or direct question\n if ($hasContext) {\n $userPrompt = $this->claude->buildRagPrompt($question, $context);\n } else {\n $userPrompt = $question;\n }\n\n \/\/ Build system prompt hierarchy: Default -> Custom -> Style\n if ($customSystemPrompt !== null && $customSystemPrompt !== '') {\n $systemPrompt = $customSystemPrompt;\n } else {\n $systemPrompt = $hasContext\n ? $this->claude->getDefaultSystemPrompt()\n : 'Du bist ein hilfreicher Assistent. Antworte auf Deutsch, präzise und hilfreich.';\n }\n\n \/\/ Append style prompt from author profile if provided\n if ($stylePrompt !== null && $stylePrompt !== '') {\n $systemPrompt .= \"\\n\\n\" . $stylePrompt;\n }\n\n $llmResponse = $this->claude->ask($userPrompt, $systemPrompt, $model, $maxTokens, $temperature);\n\n $answer = $llmResponse['text'];\n $usage = $llmResponse['usage'];\n } catch (RuntimeException $e) {\n throw new RuntimeException(\n 'Claude API request failed: ' . $e->getMessage(),\n 0,\n $e\n );\n }\n } elseif ($isOllama) {\n try {\n \/\/ Extract actual model name (remove \"ollama:\" prefix)\n $ollamaModel = substr($model, 7);\n\n \/\/ Build instruction from custom prompt and style\n $instructions = [];\n if ($customSystemPrompt !== null && $customSystemPrompt !== '') {\n $instructions[] = $customSystemPrompt;\n }\n if ($stylePrompt !== null && $stylePrompt !== '') {\n $instructions[] = $stylePrompt;\n }\n $instructionBlock = $instructions !== [] ? implode(\"\\n\\n\", $instructions) . \"\\n\\n\" : '';\n\n \/\/ Build prompt: RAG with context or direct question\n if ($hasContext) {\n $userPrompt = sprintf(\n \"%sKontext aus den Dokumenten:\\n\\n%s\\n\\n---\\n\\nFrage: %s\",\n $instructionBlock,\n $context,\n $question\n );\n } else {\n $userPrompt = $instructionBlock . $question;\n }\n\n $answer = $this->ollama->generate($userPrompt, $ollamaModel, $temperature);\n $usage = null;\n } catch (RuntimeException $e) {\n throw new RuntimeException(\n 'Ollama generation failed: ' . $e->getMessage(),\n 0,\n $e\n );\n }\n } else {\n throw new RuntimeException(\n sprintf('Unknown model \"%s\". Use claude-* or ollama:* format.', $model)\n );\n }\n\n \/\/ Step 5: Extract source information\n $sources = $this->extractSources($searchResults);\n\n \/\/ Step 6: Assemble response\n $response = [\n 'question' => $question,\n 'answer' => $answer,\n 'sources' => $sources,\n 'model' => $model,\n 'chunks_used' => count($searchResults),\n ];\n\n if ($usage !== null) {\n $response['usage'] = $usage;\n }\n\n return $response;\n }\n\n \/**\n * Builds a context string from search results.\n *\n * Concatenates the content from multiple search results into a single\n * context string, respecting a maximum character limit. Each chunk is\n * labeled with its source document title.\n *\n * @param array<int, array{id: int|string, score: float, payload: array<string, mixed>}> $searchResults Vector search results\n * @param int $maxTokens Maximum tokens to include (default: 3000)\n *\n * @return string The built context string\n *\/\n private function buildContext(array $searchResults, int $maxTokens = 3000): string\n {\n $contextParts = [];\n $totalChars = 0;\n $maxChars = $maxTokens * 4; \/\/ Approximate: 1 token ~ 4 characters\n\n foreach ($searchResults as $index => $result) {\n $payload =... [TRUNCATED-cc6b7294df6d10df]",
"structuredPatch": [
{
"oldStart": 231,
"oldLines": 17,
"newStart": 231,
"newLines": 26,
"lines": [
" * context string, respecting a maximum character limit. Each chunk is",
" * labeled with its source document title.",
" *",
"- * @param array<int, array{id: int|string, score: float, payload: array<string, mixed>}> $searchResults Vector search results",
"- * @param int $maxTokens Maximum tokens to include (default: 3000)",
"+ * Optionally prepends a semantic summary with extracted entities and taxonomy.",
" *",
"+ * @param array<int, array{id: int|string, score: float, payload: array<string, mixed>}> $searchResults Vector search results",
"+ * @param int $maxTokens Maximum tokens to include (default: 3000)",
"+ * @param string $semanticSummary Optional semantic summary to prepend",
"+ *",
" * @return string The built context string",
" *\/",
"- private function buildContext(array $searchResults, int $maxTokens = 3000): string",
"+ private function buildContext(array $searchResults, int $maxTokens = 3000, string $semanticSummary = ''): string",
" {",
" $contextParts = [];",
" $totalChars = 0;",
" $maxChars = $maxTokens * 4; \/\/ Approximate: 1 token ~ 4 characters",
" ",
"+ \/\/ Prepend semantic summary if available",
"+ if ($semanticSummary !== '') {",
"+ $contextParts[] = \"[Semantischer Kontext]\\n\" . $semanticSummary;",
"+ $totalChars += strlen($semanticSummary);",
"+ }",
"+",
" foreach ($searchResults as $index => $result) {",
" $payload = $result['payload'];",
" \/\/ Support both payload schemas: documents + dokumentation_chunks"
]
},
{
"oldStart": 253,
"oldLines": 7,
"newStart": 262,
"newLines": 15,
"lines": [
" break;",
" }",
" ",
"- $contextParts[] = sprintf('[Quelle %d: %s]%s%s', $index + 1, $docTitle, \"\\n\", $content);",
"+ \/\/ Build chunk header with optional entity info",
"+ $entities = $payload['entities'] ?? [];",
"+ $entityInfo = '';",
"+ if ($entities !== []) {",
"+ $entityNames = array_slice(array_column($entities, 'name'), 0, 3);",
"+ $entityInfo = ' | Enthält: ' . implode(', ', $entityNames);",
"+ }",
"+",
"+ $contextParts[] = sprintf('[Quelle %d: %s%s]%s%s', $index + 1, $docTitle, $entityInfo, \"\\n\", $content);",
" $totalChars += strlen($content);",
" }",
" "
]
}
],
"userModified": false,
"replaceAll": false
}
}