投标文件查重(免费)-比地招标网

概述

投标文件查重-比地招标网

Compare documents from the current request only. Treat every match as review evidence, never as a plagiarism conclusion.

Required Workflow

1. Confirm that the user is authorized to inspect every uploaded document. Accept two to ten files. Process files up to 50 MB normally. For an oversized DOCX, obtain explicit authorization before adding --allow-large-docx; this mode reads only body paragraphs and table rows locally in memory, up to 250 MB.
2. Confirm the role of every file. Identify tender, template, and reference files explicitly; unconfirmed files are treated as bids and the report records that assumption.
3. Run scripts/compare_batch.py into a new user-selected output directory. Do not write derived text beside the source documents.
4. When a tender document is present, review the Tender Baseline Exclusion section. Exact tender text shared by at least two bid documents is excluded from review-required text evidence and remains countable in the report.
5. Open summary.html and inspect extraction coverage and warnings before interpreting the result. Explain that unreadable PDF pages and failed OCR are detection blind spots.
6. Review evidence in the source documents. Do not characterize a finding as infringement, cheating, collusion, or plagiarism without an authorized human decision.
7. Deliver summary.html and evidence.json; remove any temporary workspace after delivery.

Run

python3 /Users/jinxinglin/.codex/skills/compare-tender-batch/scripts/compare_batch.py 
  --output /absolute/path/to/review-output 
  --allow-large-docx 
  --role '招标文件.docx=tender' 
  --role '企业模板.docx=template' 
  /absolute/path/to/*.docx

Pass paths explicitly when a shell wildcard could include unrelated documents. Valid roles are bid (default), tender, template, and reference.

The script accepts DOCX, TXT, Markdown, and PDF. It uses pdftotext for selectable PDF text and locally falls back to pdftoppm plus Tesseract for scanned PDFs when those tools and language data are available. Consecutive short paragraphs or rows are merged into bounded local context before comparison; any remaining short text is counted as a blind spot in the report. With explicit --allow-large-docx, an oversized DOCX is streamed in 500-block text batches without reading its media files or writing intermediate fragments; its document.xml is limited to 64 MB. The two-bid difference view suppresses document-wide unaligned replacements and shows a summary instead. It never uploads content or builds a persistent corpus. It applies exact text/near-duplicate evidence detection; it does not claim semantic paraphrase detection.

Interpretation

• Treat Review required as a non-exempt, bid-to-bid near-duplicate that needs human review. It is not a legal, procurement, or misconduct conclusion.
• Treat Reference reuse as a match involving a tender/template/reference file; verify whether the reuse is permitted.
• Treat Excluded and Common batch content as non-risk content unless a reviewer overrides the rule.
• Treat matching DOCX author/company properties as Requires authorized verification. They are redacted and never merge into the text-evidence count.
• Treat matching phone numbers, email addresses, and unified social credit code candidates as Requires authorized verification. They are redacted and never merge into the text-evidence count.
• Use the two-bid-document difference section only as a drafting view. It is unavailable for any other number of successfully extracted bid documents.
• Compare the number of extracted segments and unreadable files before using absence of findings as evidence.

Read [references/scoring-policy.md](references/scoring-policy.md) before changing thresholds or communicating conclusions. Read [references/data-handling.md](references/data-handling.md) when documents are sensitive.

Resources

• scripts/compare_batch.py: Extract files, find cross-document evidence, and render summary.html plus evidence.json.
• assets/default-exemption-rules.json: Default fixed-text exclusions; copy and adjust only with an approval trail.
• references/scoring-policy.md: Scoring, false-positive controls, and acceptance criteria.
• references/data-handling.md: Local-processing and retention rules.