Describe an image in natural language using Florence-2.

Args: image_path: Absolute or relative path to the image file (supports PNG, JPEG, SVG). detail_level: 'normal' for a brief caption, 'high' for a detailed one. model_mode: 'fast' for Florence-2 (default), 'deep' for MiniCPM-V 4.6 (better document understanding).

Returns: Dict with description, model name, and prompt used.

Extract text from an image using Florence-2 OCR.

Args: image_path: Absolute or relative path to the image file (supports PNG, JPEG, SVG). detail_level: 'normal' for plain OCR, 'high' for OCR with region info. model_mode: 'fast' for Florence-2 (default), 'deep' for MiniCPM-V 4.6 (better document understanding).

Returns: Dict with extracted text and optionally bounding regions.

Describe UI regions in a screenshot using Florence-2.

Args: image_path: Absolute or relative path to the screenshot file (supports PNG, JPEG, SVG). detail_level: 'normal' for dense region captions, 'high' for per-region descriptions. model_mode: 'fast' for Florence-2 (default), 'deep' for MiniCPM-V 4.6 (better document understanding).

Returns: Dict with detected regions (bounding boxes and labels) and model name.

Capture a screenshot and optionally describe it using Florence-2.

Args: output_path: Path to save the screenshot PNG. If None, saves to a temp file. monitor: Monitor index (0 = all monitors combined, 1 = primary, etc.). describe: If True, also run describe_screenshot on the captured image. model_mode: 'fast' for Florence-2 (default), 'deep' for MiniCPM-V 4.6 (better document understanding).

Returns: Dict with path, width, height, monitor, and optionally regions.

Extract text from an image using PaddleOCR (100+ languages, production-grade).

PaddleOCR is purpose-built for text extraction with superior accuracy and speed compared to general vision models. Best for:

• Multi-language documents (100+ languages supported)

• CPU-only servers (PP-OCRv6 Tiny is only 1.5M parameters)

• High-volume batch OCR (5.2× faster than previous versions)

Args: image_path: Absolute or relative path to the image file (supports PNG, JPEG, etc.). language: Language code - 'en' (English), 'ch' (Chinese), 'japan' (Japanese), 'korean', 'french', 'german', 'spanish', 'arabic', 'multilingual', etc. See PaddleOCR docs for full list. detail_level: 'normal' for plain text, 'high' for text with bounding boxes and confidence. use_angle_cls: If True, use angle classification to correct rotated text (default True).

Returns: Dict with extracted text, and optionally regions with bounding boxes and confidence.

Parse a document (PDF, DOCX, PPTX, HTML, MD) into structured output using Docling.

Docling is IBM's document understanding library that extracts text, tables, charts, formulas, and code blocks from multi-format documents.

Args: file_path: Absolute or relative path to the document file. Supports: PDF, DOCX, PPTX, HTML, Markdown, XLSX, Images. output_format: Output format - 'markdown' (default), 'json', 'text', or 'html'. extract_tables: If True, extract and structure tables (default True). extract_images: If True, extract embedded images (default False).

Returns: Dict with parsed content, metadata, and optionally tables/images.