Protokoll #1526

{
    "event": "PreToolUse",
    "tool_name": "Edit",
    "tool_input": {
        "file_path": "\/var\/www\/dev.campus.systemische-tools.de\/src\/Infrastructure\/AI\/ChatService.php",
        "old_string": "    \/**\n     * Extracts unique source information from search results.\n     *\n     * Collects document titles and scores from the search results,\n     * deduplicating by title to provide a clean list of sources.\n     * Optionally includes content preview if available.\n     *\n     * @param array<int, array{id: int|string, score: float, payload: array<string, mixed>}> $searchResults Vector search results\n     *\n     * @return array<int, array{title: string, score: float, content?: string}> Deduplicated source information\n     *\/\n    private function extractSources(array $searchResults): array\n    {\n        $sources = [];\n        $seen = [];\n\n        foreach ($searchResults as $result) {\n            $payload = $result['payload'];\n            $docTitle = (string) ($payload['document_title'] ?? '');\n\n            \/\/ Skip empty titles or already seen titles\n            if ($docTitle === '' || isset($seen[$docTitle])) {\n                continue;\n            }\n\n            $source = [\n                'title' => $docTitle,\n                'score' => round($result['score'], 3),\n            ];\n\n            \/\/ Optionally include content preview\n            if (isset($payload['content']) && is_string($payload['content'])) {\n                $source['content'] = $payload['content'];\n            }\n\n            $sources[] = $source;\n            $seen[$docTitle] = true;\n        }\n\n        return $sources;\n    }\n}",
        "new_string": "    \/**\n     * Searches across multiple Qdrant collections and merges results.\n     *\n     * Queries each collection with the same embedding vector, then merges\n     * and sorts results by score. Returns the top N results across all collections.\n     *\n     * @param array<int, float> $embedding   The query embedding vector\n     * @param array<string>     $collections Collections to search\n     * @param int               $limit       Maximum total results to return\n     *\n     * @return array<int, array{id: int|string, score: float, payload: array<string, mixed>}> Merged search results\n     *\/\n    private function searchMultipleCollections(array $embedding, array $collections, int $limit): array\n    {\n        $allResults = [];\n\n        foreach ($collections as $collection) {\n            try {\n                $results = $this->qdrant->search($embedding, $collection, $limit);\n\n                \/\/ Add collection name to each result's payload for reference\n                foreach ($results as &$result) {\n                    $result['payload']['_collection'] = $collection;\n                }\n\n                $allResults = array_merge($allResults, $results);\n            } catch (RuntimeException) {\n                \/\/ Skip unavailable collections, continue with others\n                continue;\n            }\n        }\n\n        \/\/ Sort by score descending\n        usort($allResults, static fn($a, $b) => $b['score'] <=> $a['score']);\n\n        \/\/ Return top N results\n        return array_slice($allResults, 0, $limit);\n    }\n\n    \/**\n     * Extracts unique source information from search results.\n     *\n     * Collects document titles and scores from the search results,\n     * deduplicating by title to provide a clean list of sources.\n     * Optionally includes content preview if available.\n     *\n     * @param array<int, array{id: int|string, score: float, payload: array<string, mixed>}> $searchResults Vector search results\n     *\n     * @return array<int, array{title: string, score: float, content?: string}> Deduplicated source information\n     *\/\n    private function extractSources(array $searchResults): array\n    {\n        $sources = [];\n        $seen = [];\n\n        foreach ($searchResults as $result) {\n            $payload = $result['payload'];\n            $docTitle = (string) ($payload['document_title'] ?? '');\n\n            \/\/ Skip empty titles or already seen titles\n            if ($docTitle === '' || isset($seen[$docTitle])) {\n                continue;\n            }\n\n            $source = [\n                'title' => $docTitle,\n                'score' => round($result['score'], 3),\n            ];\n\n            \/\/ Optionally include content preview\n            if (isset($payload['content']) && is_string($payload['content'])) {\n                $source['content'] = $payload['content'];\n            }\n\n            \/\/ Add collection info if available\n            if (isset($payload['_collection'])) {\n                $source['collection'] = $payload['_collection'];\n            }\n\n            $sources[] = $source;\n            $seen[$docTitle] = true;\n        }\n\n        return $sources;\n    }\n}"
    }
}

{
    "tool_response": {
        "filePath": "\/var\/www\/dev.campus.systemische-tools.de\/src\/Infrastructure\/AI\/ChatService.php",
        "oldString": "    \/**\n     * Extracts unique source information from search results.\n     *\n     * Collects document titles and scores from the search results,\n     * deduplicating by title to provide a clean list of sources.\n     * Optionally includes content preview if available.\n     *\n     * @param array<int, array{id: int|string, score: float, payload: array<string, mixed>}> $searchResults Vector search results\n     *\n     * @return array<int, array{title: string, score: float, content?: string}> Deduplicated source information\n     *\/\n    private function extractSources(array $searchResults): array\n    {\n        $sources = [];\n        $seen = [];\n\n        foreach ($searchResults as $result) {\n            $payload = $result['payload'];\n            $docTitle = (string) ($payload['document_title'] ?? '');\n\n            \/\/ Skip empty titles or already seen titles\n            if ($docTitle === '' || isset($seen[$docTitle])) {\n                continue;\n            }\n\n            $source = [\n                'title' => $docTitle,\n                'score' => round($result['score'], 3),\n            ];\n\n            \/\/ Optionally include content preview\n            if (isset($payload['content']) && is_string($payload['content'])) {\n                $source['content'] = $payload['content'];\n            }\n\n            $sources[] = $source;\n            $seen[$docTitle] = true;\n        }\n\n        return $sources;\n    }\n}",
        "newString": "    \/**\n     * Searches across multiple Qdrant collections and merges results.\n     *\n     * Queries each collection with the same embedding vector, then merges\n     * and sorts results by score. Returns the top N results across all collections.\n     *\n     * @param array<int, float> $embedding   The query embedding vector\n     * @param array<string>     $collections Collections to search\n     * @param int               $limit       Maximum total results to return\n     *\n     * @return array<int, array{id: int|string, score: float, payload: array<string, mixed>}> Merged search results\n     *\/\n    private function searchMultipleCollections(array $embedding, array $collections, int $limit): array\n    {\n        $allResults = [];\n\n        foreach ($collections as $collection) {\n            try {\n                $results = $this->qdrant->search($embedding, $collection, $limit);\n\n                \/\/ Add collection name to each result's payload for reference\n                foreach ($results as &$result) {\n                    $result['payload']['_collection'] = $collection;\n                }\n\n                $allResults = array_merge($allResults, $results);\n            } catch (RuntimeException) {\n                \/\/ Skip unavailable collections, continue with others\n                continue;\n            }\n        }\n\n        \/\/ Sort by score descending\n        usort($allResults, static fn($a, $b) => $b['score'] <=> $a['score']);\n\n        \/\/ Return top N results\n        return array_slice($allResults, 0, $limit);\n    }\n\n    \/**\n     * Extracts unique source information from search results.\n     *\n     * Collects document titles and scores from the search results,\n     * deduplicating by title to provide a clean list of sources.\n     * Optionally includes content preview if available.\n     *\n     * @param array<int, array{id: int|string, score: float, payload: array<string, mixed>}> $searchResults Vector search results\n     *\n     * @return array<int, array{title: string, score: float, content?: string}> Deduplicated source information\n     *\/\n    private function extractSources(array $searchResults): array\n    {\n        $sources = [];\n        $seen = [];\n\n        foreach ($searchResults as $result) {\n            $payload = $result['payload'];\n            $docTitle = (string) ($payload['document_title'] ?? '');\n\n            \/\/ Skip empty titles or already seen titles\n            if ($docTitle === '' || isset($seen[$docTitle])) {\n                continue;\n            }\n\n            $source = [\n                'title' => $docTitle,\n                'score' => round($result['score'], 3),\n            ];\n\n            \/\/ Optionally include content preview\n            if (isset($payload['content']) && is_string($payload['content'])) {\n                $source['content'] = $payload['content'];\n            }\n\n            \/\/ Add collection info if available\n            if (isset($payload['_collection'])) {\n                $source['collection'] = $payload['_collection'];\n            }\n\n            $sources[] = $source;\n            $seen[$docTitle] = true;\n        }\n\n        return $sources;\n    }\n}",
        "originalFile": "<?php\n\ndeclare(strict_types=1);\n\nnamespace Infrastructure\\AI;\n\nuse RuntimeException;\n\n\/**\n * RAG (Retrieval-Augmented Generation) Chat Service.\n *\n * Provides a complete RAG pipeline that:\n * 1. Converts questions to embeddings using Ollama\n * 2. Searches for relevant document chunks in Qdrant\n * 3. Builds context from search results\n * 4. Generates answers using Claude or Ollama\n * 5. Returns structured responses with sources and metadata\n *\n * This service orchestrates the interaction between OllamaService,\n * QdrantService, and ClaudeService to implement a production-ready\n * RAG system for document-based question answering.\n *\n * @package Infrastructure\\AI\n * @author  System Generated\n * @version 1.0.0\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 {\n                $searchResults = $this->searchMultipleCollections($queryEmbedding, $collections, $limit);\n            } catch (RuntimeException $e) {\n                throw new RuntimeException(\n                    'Vector search failed: ' . $e->getMessage(),\n                    0,\n                    $e\n                );\n            }\n\n            \/\/ Step 3: Build context from search results (if any found)\n            if ($searchResults !== []) {\n                $context = $this->buildContext($searchResults);\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 = $result['payload'];\n            $content = (string) ($payload['content'] ?? '');\n            $docTitle = (string) ($payload['document_title'] ?? 'Unbekannt');\n\n            \/\/ Check if adding this chunk would exceed ... [TRUNCATED-67a7aa5aba6fcdd3]",
        "structuredPatch": [
            {
                "oldStart": 265,
                "oldLines": 6,
                "newStart": 265,
                "newLines": 45,
                "lines": [
                    "     }",
                    " ",
                    "     \/**",
                    "+     * Searches across multiple Qdrant collections and merges results.",
                    "+     *",
                    "+     * Queries each collection with the same embedding vector, then merges",
                    "+     * and sorts results by score. Returns the top N results across all collections.",
                    "+     *",
                    "+     * @param array<int, float> $embedding   The query embedding vector",
                    "+     * @param array<string>     $collections Collections to search",
                    "+     * @param int               $limit       Maximum total results to return",
                    "+     *",
                    "+     * @return array<int, array{id: int|string, score: float, payload: array<string, mixed>}> Merged search results",
                    "+     *\/",
                    "+    private function searchMultipleCollections(array $embedding, array $collections, int $limit): array",
                    "+    {",
                    "+        $allResults = [];",
                    "+",
                    "+        foreach ($collections as $collection) {",
                    "+            try {",
                    "+                $results = $this->qdrant->search($embedding, $collection, $limit);",
                    "+",
                    "+                \/\/ Add collection name to each result's payload for reference",
                    "+                foreach ($results as &$result) {",
                    "+                    $result['payload']['_collection'] = $collection;",
                    "+                }",
                    "+",
                    "+                $allResults = array_merge($allResults, $results);",
                    "+            } catch (RuntimeException) {",
                    "+                \/\/ Skip unavailable collections, continue with others",
                    "+                continue;",
                    "+            }",
                    "+        }",
                    "+",
                    "+        \/\/ Sort by score descending",
                    "+        usort($allResults, static fn($a, $b) => $b['score'] <=> $a['score']);",
                    "+",
                    "+        \/\/ Return top N results",
                    "+        return array_slice($allResults, 0, $limit);",
                    "+    }",
                    "+",
                    "+    \/**",
                    "      * Extracts unique source information from search results.",
                    "      *",
                    "      * Collects document titles and scores from the search results,"
                ]
            },
            {
                "oldStart": 299,
                "oldLines": 6,
                "newStart": 338,
                "newLines": 11,
                "lines": [
                    "                 $source['content'] = $payload['content'];",
                    "             }",
                    " ",
                    "+            \/\/ Add collection info if available",
                    "+            if (isset($payload['_collection'])) {",
                    "+                $source['collection'] = $payload['_collection'];",
                    "+            }",
                    "+",
                    "             $sources[] = $source;",
                    "             $seen[$docTitle] = true;",
                    "         }"
                ]
            }
        ],
        "userModified": false,
        "replaceAll": false
    }
}