Since Word 2010, a text box is written twice inside mc:AlternateContent:

• an mc:Choice with the DrawingML shape (wps:txbx);
• an mc:Fallback with a VML shape (v:textbox) for older versions of Word.

Both carry the box's paragraphs. docx_xml_to_text reads every w:t it meets, so the text of every text box comes out twice.

Problem

On dev, a paragraph with a text box reading "Quarterly revenue rose 12%" comes out as:

Before the box. Quarterly revenue rose 12%
Quarterly revenue rose 12%
After the box.

The same text box in a table cell gives | Revenue | Up 12% Up 12% |.

• The box's sentence is extracted twice, so its statements and their evidence are counted twice.
• In a cell, the doubled text becomes the value the table row carries.

Reports, forms and cover pages use text boxes for titles, call-outs and side notes.

Change

docx_xml_to_text follows the markup-compatibility rule. In each mc:AlternateContent it reads the first mc:Choice and skips the later Choices and the mc:Fallback. It counts the depth of a skipped subtree to know where it ends.

A document without mc:AlternateContent reads exactly as before.

Verification

Run on macOS (arm64): Rust 1.98.1; the workspace suite against PostgreSQL 16 with pgvector 0.8.2, with Poppler 26.06.

• crates/utopia-ingest/tests/docx_text_boxes.rs builds the text box the way Word writes it: the wps shape in the Choice, the VML shape in the Fallback. It also builds a variant with a second Choice.
  • a_text_box_is_read_once: the sentence appears once, and the text before and after the box is intact.
  • a_text_box_in_a_table_cell_is_read_once: | Revenue | Up 12% |.
  • On dev with only this file added, both fail with the text above.
• cargo test --locked -p utopia-ingest: 96 unit tests and every integration test pass.
• fmt, clippy and the workspace suite:
  • cargo fmt --all --check and cargo clippy --locked --workspace --all-targets -- -D warnings are clean.
  • UTOPIA_TEST_REQUIRE_DB=1 UTOPIA_TEST_REQUIRE_PDFTOTEXT=1 cargo test --locked --workspace passes: 1,331 passed, 0 failed, 6 ignored.
  • That workspace run had this change merged with the three other utopia-ingest fixes opened alongside it: spreadsheet dates, slide line breaks and Word headings.

Notes

• A text box's text is still read where Word anchors it, inside the paragraph it belongs to.
• The Word-headings change edits the same function. The two merge without a conflict, checked locally.