Multi-Source Data Extraction and Organization

Problem

Turns text and tables scattered across images/scans, PDFs, Word, Excel/CSV, and web pages into archival Excel or Word deliverables. Typical cases include audit working papers, issue lists, remediation ledgers, and evidence tables: sources vary, but outputs must follow fixed rules—consistent headers, continuous numbering, amount in Chinese uppercase, and missing values shown as —.

How It Works

The workflow first confirms the source type and desired output, then selects an extraction path:
- Images/scans: prefer the Read tool for multimodal reading; use scripts/ocr_image.py for scripted batch runs.
- PDFs / Word: use scripts/pdf_extract.py and scripts/docx_extract.py to extract text, heading hierarchy, and tables.
- Excel / CSV: use scripts/excel_clean.py for clean, merge, dedupe, and split.
- Web pages: use scripts/web_extract.py to parse tables with pandas.read_html; fall back to extracting body lists when no tables exist.

During normalization, it follows references/formats.md: dates use YYYY年MM月DD日, amounts include numeric and Chinese uppercase forms, headers are bold and centered, and missing values are consistently set to —. The final artifacts are styled .xlsx or normalized .docx files with date-prefixed filenames for easier archiving.

Boundaries

The skill focuses on structured data entry and format normalization, not general web crawling or complex layout reconstruction. For complex table scans, multimodal reading is usually more reliable than pure OCR. easyocr downloads a Chinese model on first run, so offline environments should prefer the Read path. Web extraction depends on pages that read_html can parse; dynamically rendered or bot-protected pages are outside the stable scope.