What's new#
Tensorlake now parses Word documents (.docx) with tracked changes intact, returning structured HTML where insertions, deletions, and comments are preserved with full metadata. No more manually reviewing revision history, keep track of changes and comments programmatically.
Why it matters#
- Audit trails - Extract complete revision history for compliance and record-keeping
- Workflow automation - Route documents based on specific reviewer comments or edits
- Change analysis - Programmatically identify what was added, removed, or flagged by stakeholders
- Version control - Build diffs and approval workflows without manual document review
The problem#
Most document parsers strip tracked changes entirely. When you parse a Word document with python-docx, Pandoc, or cloud OCR APIs, you lose all revision metadata:
- python-docx: No API support for tracked changes—deletions and insertions are ignored
- Pandoc: Can preserve changes with
--track-changes=all, but output is cluttered and requires custom filters - Cloud OCR: Designed for scanned documents, not revision metadata
The underlying issue? Word stores tracked changes in complex OOXML structures (<w:del>, <w:ins>, <w:comment> nodes) that most parsers can't reconstruct.
How it works#
Tensorlake extracts tracked changes from .docx files and returns clean, structured HTML:
1from tensorlake.documentai import DocumentAI
2
3doc_ai = DocumentAI()
4
5result = doc_ai.parse_and_wait(
6 file="https://example.com/claim_report_with_tracked_changes.docx"
7)
8
9# Get HTML with tracked changes preserved
10html_content = result.pages[0].page_fragments[0].content.content
11print(html_content)Output format:
1<p>Initial damage estimates suggest total losses between $2.8M and
2<span class="comment" data-note="Michael Torres: Need to verify this upper bound">$3.4M</span>,
3<ins>based on preliminary contractor assessments,</ins> which falls within policy limits
4<del>though a complete forensic analysis is pending</del>.</p>What you get#
Tracked changes are preserved as semantic HTML:
- Deletions:
<del>removed text</del> - Insertions:
<ins>added text</ins> - Comments:
<span class="comment" data-note="comment text">highlighted text</span>
Parse with any HTML library to extract revision metadata:
1from bs4 import BeautifulSoup
2
3soup = BeautifulSoup(html_content, 'html.parser')
4
5# Extract all comments
6comments = []
7for span in soup.find_all('span', class_='comment'):
8 comments.append({
9 'text': span.get_text(strip=True),
10 'comment': span.get('data-note', '')
11 })
12
13# Extract all deletions
14deletions = [del_tag.get_text() for del_tag in soup.find_all('del')]
15for deletion in deletions:
16 print(f"Deleted: {deletion}")
17
18# Extract all insertions
19insertions = [ins_tag.get_text() for ins_tag in soup.find_all('ins')]
20for insertion in insertions:
21 print(f"Inserted: {insertion}")
22
23# Print all comments
24for comment in comments:
25 print(f"Comment: {comment['text']} - {comment['comment']}")Use cases#
Insurance claim review Extract comments from multiple adjusters and route for legal review based on flagged sections.
Contract redlining Identify all changes made by counterparties and generate change summaries automatically.
Regulatory compliance Maintain complete audit trails of document edits with author attribution and timestamps.
Collaborative editing workflows Build approval systems that trigger based on specific reviewer feedback or edit patterns.
Try it#
Colab Notebook: Tracked Changes Demo
Documentation: Parsing Documents
Parse any .docx file with tracked changes and Tensorlake automatically preserves all revision metadata.
Status#
✅ Live now in the API, SDK, and on cloud.tensorlake.ai.
Works automatically on all .docx files with tracked changes, no additional configuration needed.