Docx Toolkit

Extract text, tables, and images from .docx and legacy .doc files. Handles large documents, CJK text, and complex table structures. Includes deduplication and filtering for extracted images.

A complete toolkit for processing Microsoft Word documents (.docx and legacy .doc formats).

Capabilities

1. Text + Table Extraction (.docx)

bash

python3 {baseDir}/scripts/extract_text.py input.docx output.txt

Extracts all paragraphs and tables with structure preserved. Tables are formatted as pipe-delimited rows for easy parsing.

2. Text Extraction (Legacy .doc)

bash

python3 {baseDir}/scripts/extract_doc_text.py input.doc output.txt

Handles legacy OLE2 .doc format using olefile. Extracts Unicode text from the WordDocument stream.

3. Image Extraction (.docx)

bash

python3 {baseDir}/scripts/extract_images.py input.docx output_dir/

Extracts all embedded images with:

• Automatic deduplication (MD5 hash comparison)
• Size filtering (skips tiny icons <5KB by default)
• Sequential renaming (img_001.png, img_002.jpg, etc.)

4. Image Compression

bash

python3 {baseDir}/scripts/resize_images.py input_dir/ output_dir/ [--max-width 1024]

Batch resize/compress images for API processing (saves 50-70% on vision API costs).

Dependencies

• Python 3.6+
• python-docx — for .docx processing
• olefile — for legacy .doc processing
• Pillow — for image resizing (optional, only needed for resize script)

Install:

bash

pip3 install python-docx olefile Pillow

Use Cases

• Document analysis: Extract text for AI review/summarization
• Migration: Pull content from Word docs into other formats
• Image audit: Extract and review all embedded images
• Cost optimization: Compress images before sending to vision APIs
• Batch processing: Process multiple documents in a pipeline

Notes

• Large .doc files (>200MB) may require significant RAM for olefile processing
• Image extraction preserves original format (png/jpg/gif/etc.)
• Deduplication catches exact duplicates; near-duplicates still pass through
• CJK (Chinese/Japanese/Korean) text is fully supported in both extractors