winnerz 1.1.2

A robust Python wrapper for high-performance PDF rendering and text extraction using C++ core.

WinnerZ Python Library Documentation

Overview

The winnerz library is a robust Python wrapper designed for processing, rendering, and manipulating PDF documents. It relies on a high-performance, multi-threaded C++ core extension (winnerz_core) for intensive operations while providing seamless fallback mechanisms and caching strategies in Python.

The architecture emphasizes blistering-fast text extraction, reliability, and fault-tolerance, specifically in handling binary dependencies, dynamic core library loading, and flexible preview rendering backends (PDFium and Playwright).

Architecture

The system is divided into several conceptual layers:

1. Core Loader & Diagnostics: Handles the dynamic importing of the C++ binary (winnerz_core), including binary size verification, truncation repair, and Windows DLL directory management.
2. Document Object Model: Provides Pythonic abstractions (Document, Page) to interact with PDF files, managing resources and state safely.
3. Thread-Safe Interpreter Pipeline: A C++ native, thread-safe PDF token interpreter that leverages std::async for parallel multi-page text extraction, eliminating GIL bottlenecks.
4. Rendering Pipeline: Integrates the C++ rendering engine with fallback Python-based preview engines using pypdfium2 or playwright.
5. Geometry & Data Structures: Implements domain-specific types (Rect, Matrix, Pixmap) to standardize data flow between the C++ layer and Python runtime.

Core Loading Mechanism

The library initializes the C++ binary through _load_core(). This system provides the following safety guarantees:

• Thread Safety: Uses threading.Lock() to ensure the core is initialized exactly once.
• Retry Logic: Implements a retry loop (_CORE_IMPORT_RETRIES = 3) to mitigate transient filesystem or OS-level loading issues.
• Self-Healing: If a truncated binary is detected (e.g., due to an interrupted build or copy), _try_repair_truncated_core_binary() attempts to restore it from other valid candidate binaries in the directory.
• Diagnostic Reporting: Generates detailed error messages specifying binary ABI mismatches (e.g., GLIBC mismatches) or binary sizes to accelerate debugging.

Environment Variables

• WINNERZ_PREVIEW_BACKEND: Controls the backend used for rendering preview data when the C++ core returns placeholder data.
  • Valid values: auto (default), pdfium, playwright.
  • Resolution order for auto: Falls back from PDFium to Playwright based on availability.

Class Reference

Document

Represents a PDF document instance. It manages the lifecycle of the underlying file, featuring a lazy-loading fallback architecture: it natively opens the PDF via C++ by default and only invokes PDFium for decryption or file repair if the document is encrypted or structurally corrupted.

Constructor:

• Document(path): Resolves the path and instantly initializes the C++ core. If encryption is detected, it falls back to a temporary decryption routine.

Methods:

• __getitem__(index): Retrieves a Page object at the specified 0-based index. Supports negative indexing.
• __len__(): Returns the total number of pages in the document.
• get_all_text(): A highly optimized utility that utilizes C++ multi-threading (extract_all_text_concurrent) to extract text from all pages. It uses a dynamic hardware-concurrency batching mechanism to process pages in chunks (scaling automatically with the number of CPU cores). This entirely bypasses the Python GIL and prevents thread-exhaustion (EAGAIN) on massive 5000+ page PDFs.
• save(path): Saves the current document state. Any editing methods (like redactions, image insertions) or page merging operations (show_pdf_page) are instantly written into the PDFium memory and this method flushes them to disk with high performance.
• close(): Cleans up temporary resources, such as decrypted temporary files and in-memory editing buffers.

Page

Represents a single page within a Document.

Methods:

• get_text(mode="dict", sort=False): Extracts text content.
  • mode: Can be dict, rawdict, blocks, or text.
• get_drawings(): Extracts vector drawings and graphics, mapping them to structured dictionaries containing rect, fill, and stroke properties.
• get_pixmap(matrix=None, clip=None): Renders the page to a bitmap image (Pixmap). It attempts to render using the C++ core; if that fails or returns a placeholder, it falls back to the configured Python preview backend (PDFium or Playwright).
• redact_text(rects, output_path, min_overlap_ratio=0.0): (Legacy C++ Core) Applies text-only redaction to the specified rectangles and saves the output to a new PDF file.
• add_redact_annot(rect, fill=None): Queues a redaction annotation for a specific area (supports tuple or Rect).
• apply_redactions(images=0, graphics=0): Applies pending redactions using the In-Memory engine. Set images=1 to delete images intersecting the redaction box, or keep it 0 to strictly redact text.
• clean_contents(): Completely wipes out the vector graphics and text layer of the current page.
• insert_image(rect, stream=None): Inserts an image (from bytes) into the specified rectangle. It handles internal PDF matrix transformations automatically.
• show_pdf_page(rect, doc_src, page_idx, overlay=True): Queues a complex overlay operation. It places a page from another document (doc_src) onto the current page. The actual merge is executed efficiently during doc.save().
• rect (Property): Retrieves the bounding box of the page as a Rect.

Pixmap

Represents an uncompressed image buffer containing pixel data.

Properties:

• width, height: Dimensions in pixels.
• n: Number of channels (e.g., 4 for RGBA).
• stride: Number of bytes per row.
• samples: Raw byte array of pixel data.

Methods:

• pixel(x, y): Returns a tuple representing the pixel color at the specified coordinates.
• tobytes(fmt="raw"): Encodes the pixmap to the requested format. Supported formats include raw, rgba, png, jpg, and jpeg. Output formats other than raw require the Pillow library.

Geometry Classes

• Rect(x0, y0, x1, y1): Represents a rectangle. Provides properties for width, height, and is_empty. Overloads the & operator to compute the intersection of two rectangles.
• Matrix(sx=1.0, sy=1.0): Represents a 2D scaling matrix.

Caching Strategy

The module implements file-based caching for document instances to minimize redundant initialization and file I/O operations.

• Global Document Cache: Managed via open(path). Validates cache hits using file signature metrics (file size and modification time in nanoseconds).
• Preview Document Cache: A separate caching layer (_open_preview_pdfium_doc) strictly for the pypdfium2 rendering backend to keep the preview document context alive across multiple page renders.
• C++ Thread-Safe Font Cache: The C++ core utilizes a lock-guarded (std::recursive_mutex) internal cache for Unicode, Width, and CodeSpace maps to prevent data races during parallel text extraction.

Performance Benchmark

Thanks to the native C++ multi-threading pipeline and persistent object caching, WinnerZ outperforms established industry standards like PyMuPDF (fitz) significantly in bulk text extraction tasks.

Tested on a 185-page PDF document (2024-annual-report.pdf):

• ⏱️ PyMuPDF (fitz): ~0.44s
• 🚀 WinnerZ (get_all_text()): ~0.18s (2.5x Faster)

Dependencies

• pypdfium2: Optional but highly recommended. Used for decryption, primary preview rendering, and all In-Memory editing/redaction operations (including high-speed C-level XObject merging).
• Pillow (PIL): Optional. Required for encoding Pixmap instances to PNG/JPEG and manipulating preview images.
• playwright: Optional. Used as a secondary headless browser rendering backend if PDFium is unavailable.