Error in reading data from a table with in a table.

[sahil00yada]

Bug

In v2 version the processor is not able to get data out from a table with in a table in a consistent manner. Here is the detailed anlaysis:

Here’s a concise report on what’s happening with the missing text in your PDF-to-Markdown pipeline:

---

When converting PDFs to Markdown using Docling, some table cell texts are missing in the final output. For example, the line "Minimum 9 MOB on us performance." is present in the extracted data but is missing from the Markdown and intermediate outputs. This appears to be caused by the TableFormer model (Table Structure Model) dropping or merging certain input cell texts.

---

Steps to Reproduce

1. Run the Docling pipeline on Table.pdf
2. Observe that the text is present in the initial extraction and preprocessing stages.
3. Check the debug logs for the table structure model:
   • The text is present in the input tokens to the model.
   • The text is missing from the model’s output table cells.
4. The final Markdown output does not contain the missing line.

---

Debug Evidence

• Input tokens to Table Structure Model:
  "Minimum 9 MOB on us performance." is present.
• Output table cells from TableFormer:
  The line is missing;
• No code logic drops the line:
  The drop occurs inside the TableFormer model’s prediction step (multi_table_predict).

---

What I’ve Tried

• Verified that the text is present in all early pipeline stages.
• Added debug logging to compare input tokens and output table cells.
• Confirmed that the loss occurs inside the TableFormer model, not in Docling’s own code.

---

Expected Behavior

All input cell texts, especially short or bullet-point lines, should be preserved in the output table cells unless there is a clear, documented reason for merging or dropping them.

---

Actual Behavior

The TableFormer model drops input cell texts, resulting in missing lines in the final output.

---

Request

• Is there a way to configure TableFormer to be less aggressive in merging/dropping lines?
• Can the model be updated or retrained to preserve all input cell texts?
• Any recommended workarounds for ensuring all extracted lines are present in the output?

---

Thank you for your help!
Let me know if you need any debug logs

Docling version

v2
...

Python version

3.13.5
...

Please help us.

[dosubot]

Hi @sahil00yada! I'm Dosu and I’m helping the docling team.

There isn’t a direct TableFormer config to fully prevent merging or dropping lines, but you can experiment with a few options that may help:

• Try switching between TableFormer’s FAST and ACCURATE modes. The ACCURATE mode (now default) is designed for better handling of complex tables and may improve cell preservation in nested tables. You can set this via the pipeline options in Docling source.
• Adjust the do_cell_matching option in TableStructureOptions. Setting do_cell_matching to True matches predictions back to PDF cells, which sometimes helps preserve more text, but can also break output if PDF cells are merged across columns. Setting it to False lets the model define text cells independently source.
• The TableFormer model’s multi_table_predict function now supports a correct_overlapping_cells parameter, which controls a post-processing step for fixing overlapping cell bounding boxes. Setting this to True or False may affect whether short or nested lines are dropped or merged source.

If these don’t resolve the issue, retraining or fine-tuning TableFormer would require modifying its training data to include more examples of nested tables and short lines, then running the training pipeline. This is a non-trivial process and would require access to the model code and training scripts.

As a workaround, you could implement custom post-processing after TableFormer’s output to compare the original extracted text with the model’s output and attempt to recover any dropped lines.

Let me know if you want more details on any of these options or need help locating where to set these parameters.

To reply, just mention @dosu.

---

^{How did I do? Good | Irrelevant | Incorrect | Verbose | Hallucination | Report 🐛 | Other}