Failing table structure identification

[PeterStaar-IBM]

Table structure is failing on large tables,

anonymized.pdf

[dosubot]

Table structure identification failures on large tables are a known issue in Docling. Common causes include model limitations (especially with nested, multi-page, or irregular tables), performance bottlenecks, and strict requirements in "accurate" mode (which expects perfectly rectangular tables). You might see missing or merged cell content, dropped lines, or errors like "COULD NOT CONVERT TO RS THIS TABLE TO COMPUTE SPANS" if the table isn't rectangular or has merged/missing cells. Switching to "fast" table mode can help with irregular tables, though it may be less precise example.

Performance can degrade significantly on very large tables due to cubic time complexity in some modules, but there are ongoing efforts to improve this example. The TableFormer model may also drop or merge cell texts internally, especially in complex layouts, and there’s no config to fully prevent this, but toggling options like do_cell_matching or correct_overlapping_cells may help in some cases example.

For diagnosis, try exporting the intermediate DoclingDocument to JSON or DataFrame to see if the table structure is present before Markdown export example. If you notice sudden regressions, check if your model artifacts were silently updated—pinning them with --artifacts-path can ensure reproducibility example.

If you can share the specific error message or describe the symptoms (e.g., missing rows, merged cells, slow processing), I can help narrow down the cause and suggest more targeted workarounds.

To reply, just mention @dosu.

---

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