FAQ

Why is my table not detected?

Most likely, the reason is a false negative in the machine learning model. Unfortunately, there’s not much to be done except to search for a better model.

If you know exactly where the table is, you can skip the detection step by directly passing in the bbox. You can do this by passing a bbox (tuple of xmin, ymin, xmax, ymax) into the CroppedTable constructor.

from gmft.detectors.base import CroppedTable
table = CroppedTable(page, bbox=(x0, y0, x1, y0), confidence_score=1.0, label=0)

Afterwards, the CroppedTable can be passed into formatters as usual.

How to parse my table with merged cells?

Right now, only these types of merged cells are supported:

top headers that are merged (with hierarchical semantic information)
left headers that are merged (with hierarchical semantic information)

See the Semantic Spanning section for more information.

Likewise, tables with merged cells in the center are not supported.

Is gmft thread-safe?

Short answer: no. The default PDF backend (pypdfium2) is not thread-safe.

If you need parallelism, use multiprocessing with a spawn context and pass serializable objects (like CroppedTable) across process boundaries. Do not share live page objects across threads.

I need to tweak something (location/rotation) about a table. How do I do this?

When modifying a table’s location, bbox, or rotation, make sure to do so before passing the table to the formatter.

If you need to nudge a table, you can modify the bbox parameter.

for table in tables:
    table.bbox[1] -= 15 # moves y0 up by 15 pdf units
fts = [formatter.extract(table) for table in tables]

Likewise, you can force tables to always be unrotated (or rotated!)

from gmft.presets import ingest_pdf

tables, doc = ingest_pdf("path/to/pdf")
for table in tables:
    if isinstance(table, RotatedCroppedTable):
        table.angle = 0
    # always rotated:
    # tables[i] = RotatedCroppedTable(page=table.page, bbox=table.bbox, confidence_score=table.confidence_score, label=table.label, angle=90)
    ft = formatter.extract(table)
    # ...

ValueError: The identified boxes have significant overlap

When the table structure (what gmft calls the formatter) machine learning model produces boxes that overlap significantly, it is difficult to construct a table structure.

This error happens more frequently if you pass in large table. In this case, it may be beneficial to use a procedural algorithm instead of deep methods to parse the table. See Large Table Assumption for a comparison.

To accomplish this, try setting large_table_assumption to true.

from gmft.formatters.tatr import TATRFormatConfig, TATRTableFormatter

config = TATRFormatConfig()
config.large_table_assumption = True

formatter = TATRTableFormatter(config=config)
ft = formatter.extract(table)

What format is best for LLMs?

See the section, Passing into LLMs.

How to get tables formatted inline with text?

See the section, Passing into LLMs.

Cannot close object, library is destroyed.

-> Cannot close object, library is destroyed. This may cause a memory leak!

This warning may be an indication that you forgot to explicitly call PyPDFium2Document.close(), which is required.