Getting different output in text and text_as_html for table element

[enx-github]

Hello All,
I am getting different output in text and text_as_html for table element
Please find my element below:

{
"type": "Table",
"element_id": "1aa17421-eba3-4b6c-88fb-aadcf150366d",
"text": "Impact Name Damage 1 Low It does not have a great impact on the projects, its completion in time is not endangered 2 Medium It has an effect on the project, but only in such a way that more people have to be involved, or the submission time is postponed or there is risk of a fine/penalty 3 High If this happens, the project will not be completed on time, there is a risk of termination of the contract/project (if it\u00b4s internal) fines/penalties or withdrawal from the contract",
"metadata": {
"text_as_html": "

Impact	| Name	Damage
2	Medium	It has an effect on the project, but only in such a way that more people have to be involved, or the submission time is postponed or there is risk of a fine/penalty
3	High	If this happens, the project will not be completed on time, there is a risk of termination of the contract/project (if it\u2019s internal) fines/penalties or withdrawal from the contract

",
"filetype": "application/pdf",
"languages": [
"eng",
"deu"
],
"page_number": 2,
"parent_id": "75053e81-304c-41a8-8fc4-a1f09dc51169",
"filename": "xxx\xxx.pdf"
}

[eeshsaxena]

Late to this, but the short version is that those two fields don't come from the same place, so them disagreeing isn't really a bug in either one.

text is the text extracted for the table's region. text_as_html is whatever the table-structure model handed back. In the PDF path they're attached separately, element.metadata.text_as_html = getattr(layout_element, "text_as_html", None) in unstructured/partition/pdf.py, and the element's text has already been set by then. So one can lose a row while the other keeps it, which is exactly what happened to you: your text has the 1 Low row and your text_as_html jumps straight from the header to 2Medium.

The other tell is that your text_as_html has no <table>, <tr> or <td> in it at all, just concatenated cell text with one stray |. Well-formed output from that path is real table markup. When it comes back looking like a flat string, the structure model didn't resolve the grid, and the missing row is the same failure rather than a separate one.

Two things worth checking on your side. text_as_html is only meaningful when you partition with strategy="hi_res" and infer_table_structure=True; if the structure model never ran you can still get a populated but useless value. And metadata.table_extraction_method will tell you which path actually produced it, which is usually the fastest way to confirm what you're looking at before digging further.

Your file also lists languages: ["eng", "deu"] on a table whose visible text is English. Mixed language detection on a table region tends to mean the OCR pass was already unsure about that block, which fits a table the structure model struggled with.

If you still have the PDF, partitioning that page with infer_table_structure=True and looking at table_as_cells rather than the HTML string usually makes it obvious whether the row was dropped at detection time or at render time.