aImErYbArRaUlT
Shapes that share a visual row (a two-column body, a right-aligned value next to a label, a caption beside a figure) come out in the wrong order: the right-hand shape is emitted before the left-hand one whenever its box starts a little higher. ### Repro ```python from pptx import Presentation from pptx.util import Inches from markitdown import MarkItDown prs = Presentation() slide = prs.slides.add_slide(prs.slide_layouts[6]) # blank def box(left, top, text): tb = slide.shapes.add_textbox(Inches(left), Inches(top), Inches(4), Inches(3)) tb.text_frame.text = text box(0.5, 1.5, "LEFT column body") box(5.0, 1.4, "RIGHT column body") # right box starts 0.1" higher prs.save("repro.pptx") print(MarkItDown().convert("repro.pptx").markdown) ``` Expected: `LEFT column body` then `RIGHT column body`. Actual: `RIGHT column body` then `LEFT column body`. `PptxConverter.convert()` sorts shapes by `(x.top, x.left)`, so it orders by `top` first and has no notion of "same row". This also reproduces on the bundled `tests/test_files/test.pptx` (slide 2, a two-column text + image body). ### The fix (and the one decision I'd like your steer on) Group shapes that overlap vertically into row bands, read the bands top-to-bottom and the shapes left-to-right within a band. A full-height sidebar beside a grid is lifted onto its side so the grid stays row-major, and a full-width title/caption sharing a chart's or table's frame is read on the correct side of it. No shape is added or dropped; only the order changes. Measured against a known reading order on synthetic slides with seeded jitter (mean Spearman rank correlation, higher is better; benchmark script: https://gist.github.com/aImErYbArRaUlT/6b9ee69160719ae7bc77b9480a185cce): | layout | current `(top, left)` | fix | | --- | --- | --- | | two-column body | 0.88 | 1.00 | | same-row label / value | 0.90 | 1.00 | | grid | 0.88 | 1.00 | | grid beside a full-height sidebar | 0.38 - 0.95 | 1.00 | | full-width title over a row | 0.58 | 1.00 | | newspaper / column-major (out of scope) | 0.48 | 0.67 | It preserves the `None` / top-zero handling from #2408: a missing `top` still sorts first, and a real `top` of 0 is not treated as missing. The change is **default-on**. I made it default-on because on this benchmark it is monotone: it never scores below the current order on any layout, including the out-of-scope newspaper case, so there is no layout it makes worse. The linked PR implements it default-on; if you would rather gate it behind a kwarg until it has some mileage, say so and I will switch it. ### Real-deck check I also rendered the 40 slides where the new order differs from the current one (across docling, Unstructured and Apache POI decks), boxed and numbered each shape, and scored the two orders blind - one scorer, not knowing which image was which - before unsealing the key. Of 22 slides that have a single correct order, the fix read right on 19 and the current sort on 3; the fix beat the current sort 18 to 2. The other 18 are dense diagram / risk-matrix slides with no single correct order, where neither is better. That blind pass surfaced two slides where a first cut read worse than the current sort: a full-width title banded with the row below it (a Diageo "benefits" slide) and a full-width caption nested in a table's frame (a "this is kinda backwards" slide). Both are fixed. An intermediate bands-only version had actually reversed that nested caption (Spearman -1.00 on a synthetic version of it), which is why the final rule reads a full-width row on the correct side of the frame rather than purely left-to-right. ### Known limitations - It does not reconstruct column-major "newspaper" order (reading straight down one tall column before starting the next); it improves but does not solve that case. - It does not claim to linearize arbitrary 2D diagrams where no single correct order exists; there the content is preserved and the order is heuristic, never worse than today. - Separate, pre-existing quirk worth a follow-up: a table's `GraphicFrame` can report a stale bounding box, which may mis-band it. Out of scope here; happy to file it on its own. Validated on the benchmark above, a corpus of 66 real decks / 740 slides (no shape ever dropped or duplicated), and the blind render pass. The PR is written and tested.