Skip to content

load_artifacts splits a Word (.docx) paragraph into several lines at spell-check marks #8177

Description

@SoroushRF

🔴 Required Information

Describe the Bug:

LoadArtifactsTool breaks a paragraph of a Word document into several lines wherever Word has flagged a misspelled word.

_try_extract_docx_text in src/google/adk/tools/load_artifacts_tool.py splits word/document.xml into paragraphs with:

re.split(rf'<{p_tag}(?:[^>]*)>', xml_content)

With the usual w prefix this is <w:p(?:[^>]*)>. There is no check that the tag name ends after w:p, so it also matches every tag whose name merely starts with w:p: <w:proofErr>, <w:pPr>, <w:permStart>, <w:pgSz> and others. Word writes a <w:proofErr w:type="spellStart"/> / <w:proofErr w:type="spellEnd"/> pair around each word its spell checker flags, inside the paragraph, so the paragraph is split at both marks.

Steps to Reproduce:

  1. In Microsoft Word, type two paragraphs, with at least one misspelled word (red underline), and save as report.docx:
    • The quick brwn fox jumps ovr the lazy dog.
    • Second paragraph is spelt corectly.
  2. Run:
from google.adk.tools.load_artifacts_tool import _try_extract_docx_text

print(repr(_try_extract_docx_text(open('report.docx', 'rb').read())))

The same text reaches the model when the file is saved as an artifact and loaded through load_artifacts with a Runner.

A self-contained reproduction without Word:

import io, zipfile
from google.adk.tools.load_artifacts_tool import _try_extract_docx_text

W = 'http://schemas.openxmlformats.org/wordprocessingml/2006/main'
body = (
    '<w:p><w:r><w:t xml:space="preserve">The quick </w:t></w:r>'
    '<w:proofErr w:type="spellStart"/><w:r><w:t>brwn</w:t></w:r>'
    '<w:proofErr w:type="spellEnd"/>'
    '<w:r><w:t xml:space="preserve"> fox jumps.</w:t></w:r></w:p>'
    '<w:p><w:r><w:t>Second paragraph.</w:t></w:r></w:p>'
)
buf = io.BytesIO()
with zipfile.ZipFile(buf, 'w') as zf:
  zf.writestr(
      'word/document.xml',
      f'<w:document xmlns:w="{W}"><w:body>{body}</w:body></w:document>',
  )
print(repr(_try_extract_docx_text(buf.getvalue())))

Expected Behavior:

One line per paragraph:

'The quick brwn fox jumps ovr the lazy dog.\nSecond paragraph is spelt corectly.'

Observed Behavior:

With the file saved from Word (it contains six <w:proofErr> tags, two per flagged word):

'The quick \nbrwn\n fox jumps \novr\n the lazy dog.\nSecond paragraph is spelt \ncorectly\n.'

The self-contained reproduction prints 'The quick \nbrwn\n fox jumps.\nSecond paragraph.'.

No error is raised; the model just receives the document with extra line breaks.

Environment Details:

  • ADK Library Version (pip show google-adk): 2.14.1, also current main (d05e482)
  • Desktop OS: Windows 11 (the code path is not OS-specific)
  • Python Version (python -V): 3.12
  • Document saved with Microsoft Word 16.0 (Microsoft 365)

Model Information:

  • Are you using LiteLLM: No
  • Which model is being used: N/A, the text is wrong before it reaches any model

🟡 Optional Information

Regression:

Not known. I have not checked earlier releases.

Additional Context:

A one-line change fixes it by requiring the tag name to end after the paragraph tag: rf'<{p_tag}(?=[\s/>])[^>]*>'. I have a PR ready with a regression test.

Related but separate: #7088 changes the next line of the same function (entity decoding) and does not touch the split.

How often has this issue occurred?:

  • Always (100%), for any paragraph containing a word flagged by Word's spell checker

Prepared with AI assistance (Claude Code); I reproduced it with a file saved from Word.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

Labels

No labels
No labels

Type

Projects

No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions