{
"tool_response": {
"filePath": "\/opt\/scripts\/pipeline\/vision.py",
"oldString": "DEFAULT_VISION_MODEL = \"minicpm-v:latest\"",
"newString": "DEFAULT_VISION_MODEL = \"llama3.2-vision:11b\"",
"originalFile": "#!\/usr\/bin\/env python3\n\"\"\"\nVision analysis module for KI-System Pipeline.\nExtracts PDF pages as images and analyzes them with vision models.\n\nUsage:\n python vision.py <pdf_path> # Analyze all pages\n python vision.py <pdf_path> --page 1 # Analyze specific page\n\"\"\"\n\nimport base64\nimport os\nimport sys\nfrom pathlib import Path\n\nimport requests\n\nfrom config import OLLAMA_HOST\nfrom db import db\n\n# Default vision model (can be overridden by pipeline config)\nDEFAULT_VISION_MODEL = \"minicpm-v:latest\"\n\n# Image settings\nIMAGE_DPI = 150 # Balance between quality and size\nIMAGE_FORMAT = \"png\"\nMAX_IMAGE_SIZE_MB = 10\n\n\ndef pdf_to_images(file_path, dpi=IMAGE_DPI):\n \"\"\"\n Convert PDF pages to images.\n\n Args:\n file_path: Path to PDF file\n dpi: Resolution for image extraction\n\n Returns:\n List of dicts with page_number, image_bytes, width, height\n \"\"\"\n import fitz # PyMuPDF\n\n doc = fitz.open(file_path)\n pages = []\n\n for page_num in range(len(doc)):\n page = doc[page_num]\n\n # Render page to image\n mat = fitz.Matrix(dpi \/ 72, dpi \/ 72) # 72 is default PDF DPI\n pix = page.get_pixmap(matrix=mat)\n\n # Convert to PNG bytes\n img_bytes = pix.tobytes(IMAGE_FORMAT)\n\n pages.append(\n {\n \"page_number\": page_num + 1,\n \"image_bytes\": img_bytes,\n \"width\": pix.width,\n \"height\": pix.height,\n \"size_kb\": len(img_bytes) \/ 1024,\n }\n )\n\n doc.close()\n return pages\n\n\ndef analyze_image_ollama(image_bytes, model=DEFAULT_VISION_MODEL, prompt=None):\n \"\"\"\n Analyze an image using Ollama vision model.\n\n Args:\n image_bytes: PNG\/JPEG image as bytes\n model: Vision model name (e.g., minicpm-v:latest)\n prompt: Custom prompt (default: document analysis prompt)\n\n Returns:\n dict with analysis results\n \"\"\"\n if prompt is None:\n prompt = \"\"\"Analysiere diese Seite aus einem Schulungsdokument.\n\nBeschreibe strukturiert:\n1. **Überschriften\/Titel**: Welche Überschriften gibt es?\n2. **Hauptinhalt**: Worum geht es auf dieser Seite?\n3. **Visuelle Elemente**:\n - Gibt es Bilder\/Fotos? Was zeigen sie?\n - Gibt es Diagramme\/Charts? Was stellen sie dar?\n - Gibt es Tabellen? Was enthalten sie?\n4. **Layout**: Wie ist die Seite aufgebaut (Spalten, Boxen, etc.)?\n5. **Besonderheiten**: Gibt es Hervorhebungen, Zitate oder Callouts?\n\nAntworte auf Deutsch und sei präzise.\"\"\"\n\n # Encode image as base64\n image_base64 = base64.b64encode(image_bytes).decode(\"utf-8\")\n\n try:\n response = requests.post(\n f\"{OLLAMA_HOST}\/api\/generate\",\n json={\n \"model\": model,\n \"prompt\": prompt,\n \"images\": [image_base64],\n \"stream\": False,\n \"options\": {\"temperature\": 0.3, \"num_predict\": 2048, \"num_ctx\": 4096},\n },\n timeout=120,\n )\n response.raise_for_status()\n\n result = response.json()\n return {\n \"success\": True,\n \"analysis\": result.get(\"response\", \"\"),\n \"model\": model,\n \"eval_count\": result.get(\"eval_count\", 0),\n \"eval_duration_ms\": result.get(\"eval_duration\", 0) \/ 1_000_000,\n }\n\n except requests.exceptions.Timeout:\n return {\"success\": False, \"error\": \"Vision model timeout\"}\n except requests.exceptions.RequestException as e:\n return {\"success\": False, \"error\": str(e)}\n except Exception as e:\n return {\"success\": False, \"error\": str(e)}\n\n\ndef analyze_document(file_path, model=DEFAULT_VISION_MODEL, store_images=False, image_dir=None, progress=None):\n \"\"\"\n Analyze all pages of a PDF document.\n\n Args:\n file_path: Path to PDF file\n model: Vision model to use\n store_images: Whether to save images to disk\n image_dir: Directory for saved images (default: \/tmp\/pipeline_images)\n progress: PipelineProgress instance for live updates\n\n Returns:\n List of page analysis results\n \"\"\"\n db.log(\"INFO\", f\"Vision analysis starting: {file_path}\", f\"model={model}\")\n\n # Convert PDF to images\n pages = pdf_to_images(file_path)\n db.log(\"INFO\", f\"Extracted {len(pages)} pages from PDF\")\n\n if progress:\n progress.add_log(f\"Vision: {len(pages)} Seiten gefunden\")\n\n if image_dir is None:\n image_dir = \"\/tmp\/pipeline_images\" # noqa: S108\n\n if store_images:\n os.makedirs(image_dir, exist_ok=True)\n\n results = []\n\n for page in pages:\n page_num = page[\"page_number\"]\n db.log(\"INFO\", f\"Analyzing page {page_num}\/{len(pages)}\")\n\n # Log every page for full visibility\n if progress:\n progress.add_log(f\"Vision: Seite {page_num}\/{len(pages)} wird analysiert...\")\n\n # Optional: Save image to disk\n image_path = None\n if store_images:\n filename = f\"{Path(file_path).stem}_page_{page_num:03d}.{IMAGE_FORMAT}\"\n image_path = os.path.join(image_dir, filename)\n with open(image_path, \"wb\") as f:\n f.write(page[\"image_bytes\"])\n\n # Analyze with vision model\n analysis = analyze_image_ollama(page[\"image_bytes\"], model=model)\n\n results.append(\n {\n \"page_number\": page_num,\n \"width\": page[\"width\"],\n \"height\": page[\"height\"],\n \"size_kb\": page[\"size_kb\"],\n \"image_path\": image_path,\n \"analysis\": analysis.get(\"analysis\", \"\") if analysis[\"success\"] else None,\n \"error\": analysis.get(\"error\") if not analysis[\"success\"] else None,\n \"eval_tokens\": analysis.get(\"eval_count\", 0),\n \"eval_duration_ms\": analysis.get(\"eval_duration_ms\", 0),\n }\n )\n\n if analysis[\"success\"]:\n db.log(\"INFO\", f\"Page {page_num} analyzed: {analysis.get('eval_count', 0)} tokens\")\n else:\n db.log(\"WARNING\", f\"Page {page_num} analysis failed: {analysis.get('error')}\")\n\n return results\n\n\ndef store_page_analysis(document_id, page_results):\n \"\"\"\n Store page analysis results in database.\n\n Args:\n document_id: ID of the document in documents table\n page_results: List of page analysis results from analyze_document()\n\n Returns:\n Number of pages stored\n \"\"\"\n import json\n\n stored = 0\n\n for page in page_results:\n try:\n # Convert analysis to JSON (required by DB constraint)\n vision_json = None\n if page[\"analysis\"]:\n vision_json = json.dumps(\n {\n \"analysis\": page[\"analysis\"],\n \"tokens\": page[\"eval_tokens\"],\n \"duration_ms\": page[\"eval_duration_ms\"],\n \"width\": page[\"width\"],\n \"height\": page[\"height\"],\n },\n ensure_ascii=False,\n )\n\n db.execute(\n \"\"\"INSERT INTO document_pages\n (document_id, page_number, image_path, vision_analysis, token_count)\n VALUES (%s, %s, %s, %s, %s)\n ON DUPLICATE KEY UPDATE\n image_path = VALUES(image_path),\n vision_analysis = VALUES(vision_analysis),\n token_count = VALUES(token_count)\"\"\",\n (document_id, page[\"page_number\"], page[\"image_path\"], vision_json, page[\"eval_tokens\"]),\n )\n db.commit()\n stored += 1\n except Exception as e:\n db.log(\"ERROR\", f\"Failed to store page {page['page_number']}: {e}\")\n\n return stored\n\n\ndef run_vision_step(document_id, file_path, config=None, progress=None):\n \"\"\"\n Run vision analysis step for pipeline.\n\n Args:\n document_id: Document ID in database\n file_path: Path to PDF file\n config: Step configuration dict\n progress: PipelineProgress instance for live updates\n\n Returns:\n dict with success status and statistics\n \"\"\"\n if config is None:\n config = {}\n\n model = config.get(\"model\", DEFAULT_VISION_MODEL)\n store_images = config.get(\"store_images\", False)\n detect_images = config.get(\"detect_images\", True)\n detect_charts = config.get(\"detect_charts\", True)\n detect_tables = config.get(\"detect_tables\", True)\n\n # Build custom prompt based on config\n prompt_parts = [\"Analysiere diese Seite aus einem Schulungsdokument.\\n\\nBeschreibe strukturiert:\"]\n prompt_parts.append(\"1. **Überschriften\/Titel**: Welche Überschriften gibt es?\")\n prompt_parts.append(\"2. **Hauptinhalt**: Worum geht es auf dieser Seite?\")\n\n visual_parts = []\n if detect_images:\n visual_parts.append(\"Gibt es Bilder\/Fotos? Was zeigen sie?\")\n if detect_charts:\n visual_parts.append(\"Gibt es Diagramme\/Charts? Was stellen sie dar?\")\n if detect_tables:\n visual_parts.append(\"Gibt es Tabellen? Was enthalten sie?\")\n\n if visual_parts:\n prompt_parts.append(\"3. **Visuelle Elemente**:\\n - \" + \"\\n - \".join(visual_parts))\n\n prompt_parts.append(\"4. **Layout**: Wie ist die Seite aufgebaut?\")\n prompt_parts.append(\"5. **Schlüsselbegriffe**: Welche wichtigen Begriffe\/Konzepte werden genannt?\")\n prompt_parts.append(\"\\nAntworte auf Deutsch und sei präzise.\")\n\n # Note: prompt_parts built for future custom prompt support\n _ = \"\\n\".join(prompt_parts) # Currently unused, reserved for custom prompts\n\n try:\n # Analyze document\n results = analyze_document(file_path, model=model, store_images=store_images, progress=progress)\n\n # Store results\n stored = store_page_analysis(document_id, results)\n\n # Calculate statistics\n successful = sum(1 for r in results if r[\"analysis\"])\n total_tokens = sum(r[\"eval_tokens\"] for r in results)\n total... [TRUNCATED-525d32abe91e35d9]",
"structuredPatch": [
{
"oldStart": 19,
"oldLines": 7,
"newStart": 19,
"newLines": 7,
"lines": [
" from db import db",
" ",
" # Default vision model (can be overridden by pipeline config)",
"-DEFAULT_VISION_MODEL = \"minicpm-v:latest\"",
"+DEFAULT_VISION_MODEL = \"llama3.2-vision:11b\"",
" ",
" # Image settings",
" IMAGE_DPI = 150 # Balance between quality and size"
]
}
],
"userModified": false,
"replaceAll": false
}
}