Pdf Site Log Audit

Report format

Each report contains only two sections:

1. 重复图片问题
2. 明确的图文不符问题

Rules:

• One duplicate image group = one row only
• Do not split the same duplicate group into multiple rows
• If no issue exists in a section, write 未发现...
• Do not add recommendation sections unless the user explicitly asks
• Do not add image analysis detail sections unless the user explicitly asks
• Do not add speculative risk sections unless the user explicitly asks

Important image filtering rule

Before duplicate detection, ignore obvious non-photo assets such as repeated header/footer/template images.

Default thresholds:

• width >= 500
• height >= 350
• bytes >= 50000

If a PDF has unusual layout, adjust thresholds carefully rather than disabling filtering completely.

Script

Use the bundled script:

./scripts/analyze_pdf_site_logs.py

Single PDF

python3 scripts/analyze_pdf_site_logs.py "/path/to/file.pdf"

Whole directory, recursive

python3 scripts/analyze_pdf_site_logs.py "/path/to/folder" --recursive

With custom thresholds

python3 scripts/analyze_pdf_site_logs.py "/path/to/folder" --recursive \
  --min-width 500 --min-height 350 --min-bytes 50000

Save summary JSON

python3 scripts/analyze_pdf_site_logs.py "/path/to/folder" --recursive \
  --json-out /tmp/pdf_audit_summary.json

Run with sample re-check

python3 scripts/analyze_pdf_site_logs.py "/path/to/folder" --recursive \
  --sample-check 5 --json-out /tmp/pdf_audit_summary.json

Output behavior

• Write one Word report beside each PDF
• Report filename format: same basename + 分析报告.docx
• Print a JSON summary to stdout
• Optionally append sample_checks with random re-check results

When results look wrong

If a report claims many pages share the same image, first suspect repeated template graphics, not bad hashing.

Check:

• Repeated small images across all pages
• Very small banner-like dimensions
• Whether filtering thresholds are too low

If needed, rerun with stricter thresholds.

Validation workflow

When confidence matters:

• Re-run one suspicious PDF manually
• Re-check a sample of 3-5 PDFs after batch generation
• Compare duplicate groups and date errors against the reports

Notes

This workflow is tuned for工程/养护/施工类日志 PDF. If the user wants deeper visual mismatch inspection beyond date errors, that should be treated as a stricter second-pass review, not default batch output.