Protokoll #11600

{
    "event": "PreToolUse",
    "tool_name": "Read",
    "tool_input": {
        "file_path": "\/opt\/scripts\/pipeline\/generate.py"
    }
}

{
    "tool_response": {
        "type": "text",
        "file": {
            "filePath": "\/opt\/scripts\/pipeline\/generate.py",
            "content": "#!\/usr\/bin\/env python3\n\"\"\"\nContent Generation for KI-System\nGenerates content using RAG context, author profiles, and contracts.\n\"\"\"\n\nimport json\nimport sys\n\nsys.path.insert(0, \"\/opt\/scripts\/pipeline\")\n\nfrom config import ANTHROPIC_API_KEY, ANTHROPIC_MODEL, OLLAMA_CHAT_MODEL, OLLAMA_HOST  # noqa: I001, E402\nfrom db import db  # noqa: E402\nfrom embed import search_similar  # noqa: E402\n\n\ndef _repair_json(json_str):\n    \"\"\"\n    Attempt to repair common JSON issues from LLM output.\n\n    Fixes:\n    - Unescaped quotes in strings\n    - Missing commas between array elements\n    - Trailing commas\n    - Control characters in strings\n    \"\"\"\n    import re\n\n    # Remove control characters except newlines and tabs\n    json_str = re.sub(r\"[\\x00-\\x08\\x0b\\x0c\\x0e-\\x1f]\", \"\", json_str)\n\n    # Fix common issue: missing comma before closing bracket in arrays\n    json_str = re.sub(r'\"\\s*\\n\\s*]', '\"\\n]', json_str)\n\n    # Fix trailing commas before closing brackets\/braces\n    json_str = re.sub(r\",\\s*}\", \"}\", json_str)\n    json_str = re.sub(r\",\\s*]\", \"]\", json_str)\n\n    # Fix missing commas between array elements (string followed by string)\n    json_str = re.sub(r'\"\\s*\\n\\s*\"', '\",\\n\"', json_str)\n\n    # Fix unescaped quotes within strings (heuristic: quotes not at boundaries)\n    # This is tricky, so we do a simple fix for common patterns\n    lines = json_str.split(\"\\n\")\n    fixed_lines = []\n    for line in lines:\n        # Count quotes - if odd number and line has content, try to fix\n        quote_count = line.count('\"') - line.count('\\\\\"')\n        if quote_count % 2 != 0 and \":\" in line:\n            # Try to escape internal quotes (very basic heuristic)\n            parts = line.split(\":\", 1)\n            if len(parts) == 2:\n                key_part = parts[0]\n                value_part = parts[1]\n                # If value has odd quotes, try to balance\n                if value_part.count('\"') % 2 != 0:\n                    # Add escaped quote or remove problematic char\n                    value_part = value_part.rstrip().rstrip(\",\")\n                    if not value_part.endswith('\"'):\n                        value_part += '\"'\n                    line = key_part + \":\" + value_part\n        fixed_lines.append(line)\n\n    return \"\\n\".join(fixed_lines)\n\n\ndef get_prompt(name):\n    \"\"\"Load prompt from database by name.\"\"\"\n    cursor = db.execute(\n        \"\"\"SELECT content FROM prompts WHERE name = %s AND is_active = 1 ORDER BY version DESC LIMIT 1\"\"\",\n        (name,),\n    )\n    result = cursor.fetchone()\n    cursor.close()\n    return result[\"content\"] if result else None\n\n\ndef get_rag_context(briefing, collection=\"documents\", limit=5):\n    \"\"\"\n    Get relevant context from Qdrant based on briefing.\n    Returns list of chunks with content and metadata.\n    \"\"\"\n    results = search_similar(briefing, collection=collection, limit=limit)\n\n    context_items = []\n    for result in results:\n        context_items.append(\n            {\n                \"content\": result[\"payload\"].get(\"content\", \"\"),\n                \"source\": result[\"payload\"].get(\"document_title\", \"Unknown\"),\n                \"score\": round(result[\"score\"], 4),\n            }\n        )\n\n    return context_items\n\n\ndef get_config_item(item_id, item_type):\n    \"\"\"Load configuration item from content_config table.\"\"\"\n    if not item_id:\n        return None\n\n    cursor = db.execute(\n        \"SELECT name, content FROM content_config WHERE id = %s AND type = %s AND status = 'active'\",\n        (item_id, item_type),\n    )\n    result = cursor.fetchone()\n    cursor.close()\n\n    if result:\n        config = json.loads(result[\"content\"]) if isinstance(result[\"content\"], str) else result[\"content\"]\n        return {\"name\": result[\"name\"], \"config\": config}\n    return None\n\n\ndef get_author_profile(profile_id):\n    \"\"\"Load author profile from database.\"\"\"\n    return get_config_item(profile_id, \"author_profile\")\n\n\ndef get_contract(contract_id):\n    \"\"\"Load content contract from database.\"\"\"\n    return get_config_item(contract_id, \"contract\")\n\n\ndef get_structure(structure_id):\n    \"\"\"Load content structure from database.\"\"\"\n    result = get_config_item(structure_id, \"structure\")\n    if result:\n        # Structure has additional 'type' field in config\n        result[\"type\"] = result[\"config\"].get(\"type\", \"article\")\n    return result\n\n\ndef get_order(order_id):\n    \"\"\"Load content order with all related data.\"\"\"\n    cursor = db.execute(\n        \"\"\"SELECT co.*,\n                  ap.name as profile_name, ap.content as profile_config,\n                  cc.name as contract_name, cc.content as contract_config,\n                  cs.name as structure_name, cs.content as structure_config\n           FROM content_orders co\n           LEFT JOIN content_config ap ON co.author_profile_id = ap.id AND ap.type = 'author_profile'\n           LEFT JOIN content_config cc ON co.contract_id = cc.id AND cc.type = 'contract'\n           LEFT JOIN content_config cs ON co.structure_id = cs.id AND cs.type = 'structure'\n           WHERE co.id = %s\"\"\",\n        (order_id,),\n    )\n    result = cursor.fetchone()\n    cursor.close()\n    return result\n\n\ndef build_generation_prompt(briefing, context, profile, contract, structure=None):\n    \"\"\"Build the content generation prompt.\"\"\"\n\n    # Format context\n    context_text = \"\"\n    for i, ctx in enumerate(context, 1):\n        context_text += f\"\\n[Quelle {i}: {ctx['source']}]\\n{ctx['content']}\\n\"\n\n    # Build rich profile instructions\n    profile_text = \"\"\n    if profile:\n        config = profile.get(\"config\", {})\n        autorenprofil = config.get(\"autorenprofil\", config)  # Support both structures\n\n        # Extract voice\/stimme\n        stimme = autorenprofil.get(\"stimme\", {})\n        stimme_text = \"\"\n        if stimme:\n            stimme_text = f\"\"\"\n### Stimme\/Tonalität:\n- Ton: {stimme.get(\"ton\", \"neutral\")}\n- Perspektive: {stimme.get(\"perspektive\", \"neutral\")}\n- Komplexität: {stimme.get(\"komplexitaet\", \"mittel\")}\n- Autorität: {stimme.get(\"autoritaet\", \"Experte\")}\"\"\"\n\n        # Extract style\/stil\n        stil = autorenprofil.get(\"stil\", {})\n        stil_text = \"\"\n        if stil:\n            fachsprache_beispiele = stil.get(\"fachsprache_beispiele\", [])\n            fachsprache_str = \", \".join(fachsprache_beispiele[:5]) if fachsprache_beispiele else \"\"\n            stil_text = f\"\"\"\n### Stil:\n- Fachsprache: {\"Ja\" if stil.get(\"fachsprache\", False) else \"Nein\"}\n- Fachbegriffe: {fachsprache_str}\n- Satzlänge: {stil.get(\"satzlaenge\", \"mittel\")}\n- Absatzstruktur: {stil.get(\"absatzstruktur\", \"klar gegliedert\")}\"\"\"\n\n        # Extract language patterns\/sprachliche_muster\n        muster = autorenprofil.get(\"sprachliche_muster\", {})\n        muster_text = \"\"\n        if muster:\n            phrasen = muster.get(\"phrasen\", [])\n            phrasen_str = \", \".join([f'\"{p}\"' for p in phrasen[:4]]) if phrasen else \"\"\n            uebergaenge = muster.get(\"uebergaenge\", [])\n            uebergaenge_str = \", \".join([f'\"{u}\"' for u in uebergaenge[:3]]) if uebergaenge else \"\"\n            muster_text = f\"\"\"\n### Sprachliche Muster:\n- Typische Phrasen: {phrasen_str}\n- Übergangsformulierungen: {uebergaenge_str}\"\"\"\n\n        # Extract taboos\/tabus\n        tabus = autorenprofil.get(\"tabus\", [])\n        tabus_text = \"\"\n        if tabus:\n            tabus_str = \", \".join(tabus[:5])\n            tabus_text = f\"\"\"\n### Zu vermeiden (Tabus):\n{tabus_str}\"\"\"\n\n        # Extract core messages\/kernbotschaften\n        kernbotschaften = autorenprofil.get(\"kernbotschaften\", [])\n        kern_text = \"\"\n        if kernbotschaften:\n            kern_str = \"\\n\".join([f\"- {k}\" for k in kernbotschaften[:4]])\n            kern_text = f\"\"\"\n### Kernbotschaften (implizit vermitteln):\n{kern_str}\"\"\"\n\n        # Extract output format (HTML for web profiles)\n        output_format = autorenprofil.get(\"output_format\", {})\n        format_text = \"\"\n        if output_format and output_format.get(\"typ\") == \"html\":\n            erlaubte = output_format.get(\"erlaubte_elemente\", {})\n            verboten = output_format.get(\"verbotene_elemente\", [])\n            regeln = output_format.get(\"struktur_regeln\", [])\n\n            erlaubte_str = []\n            for _kategorie, elemente in erlaubte.items():\n                erlaubte_str.extend(elemente[:3])\n\n            format_text = f\"\"\"\n### Output-Format: HTML\n**Erlaubte Elemente:**\n{\", \".join(erlaubte_str[:10])}\n\n**Struktur-Regeln:**\n- {chr(10).join(\"- \" + r for r in regeln[:4]) if regeln else \"Semantisches HTML verwenden\"}\n\n**Verboten:** {\", \".join(verboten[:4]) if verboten else \"div, span, br, style\"}\"\"\"\n\n        # Combine all profile sections\n        profile_text = f\"\"\"\n## Autorenprofil: {profile.get(\"name\", \"Standard\")}\n{stimme_text}\n{stil_text}\n{muster_text}\n{tabus_text}\n{kern_text}\n{format_text}\n\"\"\"\n\n    # Build contract requirements\n    contract_text = \"\"\n    if contract:\n        config = contract.get(\"config\", {})\n        req = config.get(\"requirements\", {})\n        contract_text = f\"\"\"\nContract: {contract.get(\"name\", \"Standard\")}\n- Wortanzahl: {req.get(\"min_words\", 500)} - {req.get(\"max_words\", 5000)} Wörter\n- Struktur-Validierung: {req.get(\"structure_validation\", True)}\n\"\"\"\n\n    # Build structure instructions\n    structure_text = \"\"\n    if structure:\n        config = structure.get(\"config\", {})\n        structure_text = f\"\"\"\nStruktur-Template: {structure.get(\"name\", \"\")}\n- Typ: {structure.get(\"type\", \"\")}\n- Abschnitte: {json.dumps(config.get(\"sections\", []), ensure_ascii=False)}\n\"\"\"\n\n    # Determine if HTML output is required\n    html_output = False\n    if profile:\n        config = profile.get(\"config\", {})\n        autorenprofil = config.get(\"autorenprofil\", config)\n        output_format = autorenprofil.get(\"output_format\", {})\n        html_output = output_format.get(\"typ\") == \"html\"\n\n    format_instruction = \"\"\n    if html_output:\n        format_instruction = \"\"\"7. **WICHTIG: Ausgabe als wohlformatiertes HTML**\n   - Verwende nur: h1, h2, h3, h4, p, ul, ol, li, strong, a, section, table, tr, td\n   - Jeder h2-Abschnitt mu... [TRUNCATED-b1a598d104e1d301]",
            "numLines": 863,
            "startLine": 1,
            "totalLines": 863
        }
    }
}