Protokoll #5355

{
    "event": "PreToolUse",
    "tool_name": "Edit",
    "tool_input": {
        "file_path": "\/opt\/scripts\/pipeline\/knowledge.py",
        "old_string": "    if len(sys.argv) < 2:\n        print(\"Verwendung: python knowledge.py <document_id> [provider] [model]\")\n        print(\"  provider: ollama (default) oder anthropic\")\n        print(\"  model: mistral (ollama) oder claude-3-haiku-20240307 (anthropic)\")\n        sys.exit(1)",
        "new_string": "    if len(sys.argv) < 2:\n        print(\"Verwendung: python knowledge.py <document_id> [provider] [model]\")\n        print(\"  provider: ollama (default) oder anthropic\")\n        print(\"  model: gemma3:27b-it-qat (ollama) oder claude-3-haiku-20240307 (anthropic)\")\n        sys.exit(1)"
    }
}

{
    "tool_response": {
        "filePath": "\/opt\/scripts\/pipeline\/knowledge.py",
        "oldString": "    if len(sys.argv) < 2:\n        print(\"Verwendung: python knowledge.py <document_id> [provider] [model]\")\n        print(\"  provider: ollama (default) oder anthropic\")\n        print(\"  model: mistral (ollama) oder claude-3-haiku-20240307 (anthropic)\")\n        sys.exit(1)",
        "newString": "    if len(sys.argv) < 2:\n        print(\"Verwendung: python knowledge.py <document_id> [provider] [model]\")\n        print(\"  provider: ollama (default) oder anthropic\")\n        print(\"  model: gemma3:27b-it-qat (ollama) oder claude-3-haiku-20240307 (anthropic)\")\n        sys.exit(1)",
        "originalFile": "\"\"\"\nModulares Wissensextraktions-System für 3-Ebenen-Analyse.\n\nEbenen:\n1. Seite (page): Detailanalyse pro Einzelseite\n2. Abschnitt (section): Aggregation pro Kapitel\/Unterkapitel\n3. Dokument (document): Konsolidierte Gesamtsicht\n\nWissenstypen:\n- Entitäten: Personen, Konzepte, Methoden, Organisationen\n- Semantik: Bedeutung, Definition, Referenzen\n- Ontologie: Wechselwirkungen zwischen Entitäten\n- Taxonomie: Hierarchische Einordnung\n\nAutor: KI-System Pipeline\n\"\"\"\n\nimport json\nimport re\nimport time\nfrom dataclasses import dataclass\nfrom enum import Enum\n\nimport requests\n\nfrom config import ANTHROPIC_API_KEY, OLLAMA_HOST\nfrom db import db\n\n\nclass KnowledgeLevel(Enum):\n    \"\"\"Ebene der Wissensextraktion.\"\"\"\n\n    PAGE = \"page\"\n    SECTION = \"section\"\n    DOCUMENT = \"document\"\n\n\nclass KnowledgeType(Enum):\n    \"\"\"Typ des extrahierten Wissens.\"\"\"\n\n    ENTITY = \"entity\"\n    SEMANTIC = \"semantic\"\n    ONTOLOGY = \"ontology\"\n    TAXONOMY = \"taxonomy\"\n\n\n@dataclass\nclass ModelConfig:\n    \"\"\"Konfiguration für LLM-Modell.\"\"\"\n\n    provider: str  # 'ollama' oder 'anthropic'\n    model_name: str\n    temperature: float = 0.3\n    max_tokens: int = 2000\n\n\n# Standard-Modellkonfigurationen\nDEFAULT_MODELS = {\n    \"ollama\": ModelConfig(\"ollama\", \"gemma3:27b-it-qat\"),\n    \"anthropic\": ModelConfig(\"anthropic\", \"claude-3-haiku-20240307\"),\n    \"anthropic_opus\": ModelConfig(\"anthropic\", \"claude-opus-4-5-20251101\"),\n}\n\n\nclass KnowledgeExtractor:\n    \"\"\"\n    Modulare Wissensextraktion mit Datenbankabgleich.\n\n    Verwendung:\n        extractor = KnowledgeExtractor(model_config)\n\n        # Pro Seite\n        entities = extractor.extract_entities(text, KnowledgeLevel.PAGE, page_id)\n        semantics = extractor.extract_semantics(entities, text, KnowledgeLevel.PAGE, page_id)\n        ontology = extractor.extract_ontology(entities, text, KnowledgeLevel.PAGE, page_id)\n        taxonomy = extractor.extract_taxonomy(entities, text, KnowledgeLevel.PAGE, page_id)\n    \"\"\"\n\n    def __init__(self, model_config: ModelConfig | None = None):\n        \"\"\"Initialisiere Extractor mit Modellkonfiguration.\"\"\"\n        self.model = model_config or DEFAULT_MODELS[\"ollama\"]\n        self.anthropic_client = None\n\n        if self.model.provider == \"anthropic\":\n            self._init_anthropic()\n\n    def _init_anthropic(self):\n        \"\"\"Initialisiere Anthropic Client.\"\"\"\n        try:\n            import anthropic\n\n            if ANTHROPIC_API_KEY:\n                self.anthropic_client = anthropic.Anthropic(api_key=ANTHROPIC_API_KEY)\n        except ImportError:\n            db.log(\"WARNING\", \"Anthropic SDK nicht installiert, fallback zu Ollama\")\n            self.model = DEFAULT_MODELS[\"ollama\"]\n\n    def _call_llm(self, prompt: str, json_output: bool = True) -> str:\n        \"\"\"Rufe LLM auf und gib Antwort zurück.\"\"\"\n        start_time = time.time()\n\n        try:\n            if self.model.provider == \"anthropic\" and self.anthropic_client:\n                response = self.anthropic_client.messages.create(\n                    model=self.model.model_name,\n                    max_tokens=self.model.max_tokens,\n                    temperature=self.model.temperature,\n                    messages=[{\"role\": \"user\", \"content\": prompt}],\n                )\n                result = response.content[0].text\n                tokens_in = response.usage.input_tokens\n                tokens_out = response.usage.output_tokens\n            else:\n                # Ollama\n                payload = {\n                    \"model\": self.model.model_name,\n                    \"prompt\": prompt,\n                    \"stream\": False,\n                    \"options\": {\"temperature\": self.model.temperature},\n                }\n                if json_output:\n                    payload[\"format\"] = \"json\"\n\n                resp = requests.post(f\"{OLLAMA_HOST}\/api\/generate\", json=payload, timeout=120)\n                resp.raise_for_status()\n                data = resp.json()\n                result = data.get(\"response\", \"\")\n                tokens_in = data.get(\"prompt_eval_count\", 0)\n                tokens_out = data.get(\"eval_count\", 0)\n\n            duration_ms = int((time.time() - start_time) * 1000)\n\n            # Protokolliere LLM-Aufruf\n            db.log_to_protokoll(\n                client_name=\"pipeline-knowledge\",\n                request=prompt[:500],\n                response=result[:500],\n                model_name=f\"{self.model.provider}:{self.model.model_name}\",\n                tokens_input=tokens_in,\n                tokens_output=tokens_out,\n                duration_ms=duration_ms,\n                status=\"completed\",\n            )\n\n            return result\n\n        except Exception as e:\n            db.log(\"ERROR\", f\"LLM-Aufruf fehlgeschlagen: {e}\")\n            return \"{}\"\n\n    def _parse_json(self, text: str) -> dict:\n        \"\"\"Extrahiere JSON aus LLM-Antwort.\"\"\"\n        try:\n            # Versuche direkt zu parsen\n            return json.loads(text)\n        except json.JSONDecodeError:\n            # Suche nach JSON-Block\n            match = re.search(r\"\\{[\\s\\S]*\\}\", text)\n            if match:\n                try:\n                    return json.loads(match.group())\n                except json.JSONDecodeError:\n                    pass\n        return {}\n\n    # =========================================================================\n    # ENTITÄTEN\n    # =========================================================================\n\n    def extract_entities(self, text: str, level: KnowledgeLevel, source_id: int) -> list[dict]:\n        \"\"\"\n        Extrahiere Entitäten aus Text.\n\n        Args:\n            text: Eingabetext\n            level: Ebene (PAGE, SECTION, DOCUMENT)\n            source_id: ID der Quelle (page_id, section_id, document_id)\n\n        Returns:\n            Liste von Entitäten mit DB-IDs\n        \"\"\"\n        prompt = f\"\"\"Analysiere den folgenden deutschen Text und extrahiere alle wichtigen Entitäten.\n\nKategorien:\n- PERSON: Namen von Personen, Autoren, Therapeuten\n- ORGANIZATION: Firmen, Institute, Verbände\n- CONCEPT: Fachbegriffe, Theorien, Modelle\n- METHOD: Methoden, Techniken, Verfahren\n- TOOL: Werkzeuge, Instrumente, Materialien\n- LOCATION: Orte, Länder, Regionen\n- EVENT: Ereignisse, Konferenzen\n\nAntworte NUR als JSON:\n{{\"entities\": [\n    {{\"name\": \"...\", \"type\": \"CONCEPT\", \"context\": \"kurzer Kontext\", \"importance\": 0.0-1.0}}\n]}}\n\nText ({level.value}-Ebene):\n{text[:4000]}\"\"\"\n\n        result = self._call_llm(prompt)\n        data = self._parse_json(result)\n        entities = data.get(\"entities\", [])\n\n        # Speichere und gleiche mit DB ab\n        stored_entities = []\n        for entity in entities:\n            stored = self._store_entity(entity, level, source_id)\n            if stored:\n                stored_entities.append(stored)\n\n        # Speichere in page_knowledge\/section_knowledge\/document_knowledge\n        self._store_knowledge(\n            level,\n            source_id,\n            KnowledgeType.ENTITY,\n            {\"entities\": [e[\"name\"] for e in stored_entities], \"count\": len(stored_entities)},\n        )\n\n        return stored_entities\n\n    def _store_entity(self, entity: dict, level: KnowledgeLevel, source_id: int) -> dict | None:\n        \"\"\"Speichere Entität in DB mit Duplikatprüfung.\"\"\"\n        try:\n            name = entity.get(\"name\", \"\").strip()\n            entity_type = entity.get(\"type\", \"OTHER\").upper()\n            context = entity.get(\"context\", \"\")\n\n            if not name:\n                return None\n\n            # Prüfe ob Entität existiert (case-insensitive)\n            cursor = db.execute(\n                \"\"\"SELECT id, name, type, description\n                   FROM entities\n                   WHERE LOWER(name) = LOWER(%s) OR LOWER(canonical_name) = LOWER(%s)\n                   LIMIT 1\"\"\",\n                (name, name),\n            )\n            existing = cursor.fetchone()\n            cursor.close()\n\n            if existing:\n                entity_id = existing[\"id\"]\n                # Entität existiert - verwende bestehende\n                db.log(\"DEBUG\", f\"Entität '{name}' existiert bereits (ID: {entity_id})\")\n            else:\n                # Neue Entität anlegen\n                cursor = db.execute(\n                    \"\"\"INSERT INTO entities (name, type, canonical_name, created_at)\n                       VALUES (%s, %s, %s, NOW())\"\"\",\n                    (name, entity_type, name.lower()),\n                )\n                db.commit()\n                entity_id = cursor.lastrowid\n                cursor.close()\n                db.log(\"INFO\", f\"Neue Entität angelegt: '{name}' (ID: {entity_id})\")\n\n            return {\"id\": entity_id, \"name\": name, \"type\": entity_type, \"context\": context, \"is_new\": existing is None}\n\n        except Exception as e:\n            db.log(\"ERROR\", f\"Fehler beim Speichern der Entität: {e}\")\n            return None\n\n    # =========================================================================\n    # SEMANTIK (Bedeutung\/Definition\/Referenz)\n    # =========================================================================\n\n    def extract_semantics(self, entities: list[dict], text: str, level: KnowledgeLevel, source_id: int) -> list[dict]:\n        \"\"\"\n        Extrahiere Semantik (Bedeutung\/Definition) für Entitäten.\n\n        Args:\n            entities: Liste der extrahierten Entitäten\n            text: Ursprungstext für Kontext\n            level: Ebene\n            source_id: Quell-ID\n\n        Returns:\n            Liste von Semantik-Einträgen\n        \"\"\"\n        if not entities:\n            return []\n\n        entity_names = [e[\"name\"] for e in entities[:15]]\n\n        prompt = f\"\"\"Für die folgenden Entitäten aus dem Text, extrahiere die Bedeutung\/Definition.\n\nEntitäten: {\", \".join(entity_names)}\n\nFür jede Entität gib an:\n- definition: Kurze Definition basierend auf dem Text\n- context: In welchem Kontext wird sie verwendet\n- references: Bezüge zu anderen Konzepten (falls erkennbar)\n\nAntworte NUR als JSON:\n{{\"semantics\": [\n    {{\"entity\": \"...\", \"definiti... [TRUNCATED-22f8f5972456e30e]",
        "structuredPatch": [
            {
                "oldStart": 878,
                "oldLines": 7,
                "newStart": 878,
                "newLines": 7,
                "lines": [
                    "     if len(sys.argv) < 2:",
                    "         print(\"Verwendung: python knowledge.py <document_id> [provider] [model]\")",
                    "         print(\"  provider: ollama (default) oder anthropic\")",
                    "-        print(\"  model: mistral (ollama) oder claude-3-haiku-20240307 (anthropic)\")",
                    "+        print(\"  model: gemma3:27b-it-qat (ollama) oder claude-3-haiku-20240307 (anthropic)\")",
                    "         sys.exit(1)",
                    " ",
                    "     doc_id = int(sys.argv[1])"
                ]
            }
        ],
        "userModified": false,
        "replaceAll": false
    }
}