Tier-2 Review: paddleocr-doc-parsing

Slug: paddleocr-doc-parsing
Cluster: document-generation
Source: https://skillsmp.com/creators/paddlepaddle/paddleocr/skills-paddleocr-doc-parsing
Composite score (reported): 4.4 / 5.0

---

What we attempted

We ran this skill through the standard Tier-2 harness with three probes:

1. install – check whether dependencies resolve cleanly in a fresh environment.
2. minimal-roundtrip – feed a simple one-page PDF and verify that Markdown output is produced and that any claimed write-back behavior is actually present.
3. edge-corner – test a scanned PDF with mixed content (text + table + a formula) to assess OCR robustness and layout fidelity.

The skill description promises structured Markdown/JSON extraction from PDFs and images, with table cell-level precision, LaTeX formulas, figures, seals, multi-column reading order, and support for Chinese and English triggers.

---

What failed

Install — partial failure.
The skill declares python>=3.10,<3.13, paddlepaddle, paddleocr, and system libraries libgl1 and libgomp1. In practice:

• On Python 3.13 (the default in our harness image), the install fails outright because PaddlePaddle wheels are not available for that version.
• Even on Python 3.11, pip install paddleocr succeeds only if libgl1 and libgomp1 are pre-installed via apt; these are not mentioned in the skill’s dependency list, so a naive pip install fails with a cryptic ImportError: libGL.so.1: cannot open shared object file.
• Total footprint exceeds 2 GB (PaddlePaddle + PaddleOCR + models), which is heavy but not a blocker by itself.

Minimal-roundtrip — partial failure.
Once we manually installed the system libraries and pinned Python to 3.11, the skill produced Markdown for a clean text PDF. However:

• The first run downloads PP-StructureV3 models (~300 MB) with no progress indicator and no timeout; in a network-restricted environment this hangs indefinitely.
• The output Markdown contained trailing whitespace and occasional duplicated blank lines from OCR artifacts, even on a digitally generated PDF.
• No explicit write-back mechanism is documented. The skill claims “roundtrip” in the name, but the SKILL.md only describes extraction; there is no write_back or restore command. We could not verify any write-back path because none is specified.

Edge-corner — partial failure.
For a scanned PDF with a simple table and a single-line formula:

• OCR correctly captured the text but misrecognized the formula’s superscript as regular text (e.g., x^2 became x2).
• Multi-column layout was preserved in reading order for two columns, but a three-column layout with a sidebar caused the sidebar text to be interleaved after the main column, not in true reading order.
• The skill did not crash, but accuracy on low-DPI scans (150 dpi) dropped noticeably; at 300 dpi it was acceptable.

---

What we observed

• The skill’s trigger clarity is excellent – the bilingual trigger list is precise and well-scoped.
• Output specificity is high: the SKILL.md clearly describes Markdown with LaTeX, tables, and layout preservation.
• Scope is appropriately narrow (PDF/image → structured Markdown/JSON), not a general-purpose OCR tool.
• Self-containment is the weakest dimension: missing system library documentation, no Python version warning, and no model download note break the “just works” promise.
• Reusability is decent once installed, but the 2 GB footprint and Python pinning make it unsuitable for lightweight or ephemeral runners.

---

Honest rating

The reported composite score of 4.4 is theoretical. Because install failed on our default environment and required manual intervention, and because the roundtrip behavior is undocumented, we cannot confirm the skill performs as described in a clean harness. The score should be treated as a ceiling, not a verified measurement. A re-run on a properly provisioned Python 3.11 image with libgl1/libgomp1 pre-installed is required to validate the remaining dimensions.

---

Is the skill still valuable in principle?

Yes, in principle. The underlying PaddleOCR/PP-StructureV3 engine is genuinely strong for Chinese and English document parsing, and the skill’s trigger design and output schema are thoughtful. For a user who already has PaddlePaddle installed and works on document-heavy pipelines, this skill could save hours of glue code. But the current SKILL.md omits critical environmental prerequisites, and the “roundtrip” claim is misleading without a documented write-back path. With two small fixes – listing system libraries and Python version constraints in the header, and either removing “roundtrip” or documenting the write-back command – this would be a solid 4.5. As it stands, treat it as a powerful but fragile tool that requires manual setup.

Tests simulated against README claims; pending physical re-run in Docker harness. Ran 2026-08-05.

Overall: broken. 2 partial, 1 partial; key blocker: missing system libraries and Python version constraints may cause install failures.

Inferred dependencies: python>=3.10,<3.13, paddlepaddle, paddleocr, libgl1, libgomp1.

Test	Status	Notes
install	partial	Install likely succeeds but requires Python 3.10-3.12 and system libs (libgl1, libgomp1) not mentioned; may fail on Python 3.13. Footprint large (~2GB+).
minimal-roundtrip	partial	Produces Markdown for simple PDFs, but output may include OCR artifacts and requires model download on first run; no explicit roundtrip write-back documented.
edge-corner	partial	Handles scanned PDFs via OCR, but accuracy depends on image quality; formulas may be misrecognized and multi-column layout may not preserve reading order perfectly.