megaparse is currently unusable due to tightly coupled and broken dependencies

Summary

The installation of megaparse fails in real-world environments due to:

• Strict version locks on unavailable dependencies
• Runtime crashes from unnecessary GUI-related requirements (like OpenCV)
• Missing or undocumented system-level dependencies (despite a partial mention in README)
• Lack of a working version matrix or pinned compatible configuration
• And — critically — a lack of working examples or sufficient documentation, turning setup into frustrating guesswork

---

1. unstructured==0.15.0 is required — but it's no longer on PyPI

• megaparse==0.0.55 depends strictly on unstructured==0.15.0
• That version has been yanked from PyPI
• Installing megaparse now results in a resolution error or missing import crash

❗ There’s no published workaround or updated version that relaxes this constraint.

---

2. cv2 is included (via onnxtr?) and breaks Docker/headless environments

Even though the library doesn’t directly rely on GUI:

• OpenCV is included implicitly
• Runtime fails with: ImportError: libGL.so.1: cannot open shared object file: No such file or directory
• This can be fixed with opencv-python-headless, but that’s not exposed as a selectable install path.

4. No installable version matrix exists

Because of strict == version locks, there's no way to:

• Downgrade to a working set of packages
• Use a “core” subset of megaparse (e.g. no PDF, no OCR)
• Select only non-broken components (e.g. skip unstructured/PDF parsing)

5. Lack of documentation and examples makes setup exhausting

There are no:

• Minimal working examples
• End-to-end examples with realistic files
• Clear instructions on configuring runtime dependencies (e.g. OpenAI keys, OCR paths)

As a result, getting megaparse to work involves blind debugging and reverse-engineering library internals.

I’ve personally wasted an enormous amount of time trying to get this running, only to hit a dead end with ImportError, ResolutionImpossible, and broken package graphs.

Recommendations

• Replace hard == locks with version ranges
• Use extras_require to isolate PDF/image/ONNX features
• Vendor known working unstructured code if necessary
• Provide a working Dockerfile or requirements.lock setup for reproducibility
• Add runtime checks for system dependencies (e.g. shutil.which('tesseract'))
• Publish supported versions of each transitive dependency (pdfminer.six, pdfplumber, etc.)
• Include minimal working examples to validate installation and usage

Thanks for the work on this project, the functionality looks promising, but the current state blocks usage entirely.