dsh-md-convert

Drop in a Word file, an Excel workbook, a PowerPoint deck, or even a scanned PDF, and get back Markdown with heading levels, lists, pipe tables, and paragraph order all intact. That is the problem dsh-md-convert tackles: when you are building a RAG knowledge base, preparing documents for an LLM pipeline, or archiving a mixed-document inbox, losing structural fidelity halfway through defeats the purpose. It sits on the MarkItDown engine and treats structure-preserving conversion as a hard requirement, not a happy coincidence.

The real differentiator is how it handles scanned PDFs. Instead of a brute-force whole-page OCR pass, it runs a modular routing pipeline: PP-DocLayout-L analyses the page layout first, then dispatches each region to the right specialist - RapidOCR for text, SLANet plus RT-DETR for tables, FormulaNet for equations, and a comment annotation for stamps. Every model is lightweight and runs on CPU. Models are downloaded once into a local cache and all subsequent runs are fully offline with no network check. Legacy formats like .doc, .xls, and .ppt are auto-detected and saved as modern equivalents via WPS, Office COM, or LibreOffice before hitting MarkItDown, so platform quirks stay invisible to the user.

It ships two entry points: a batch-friendly CLI and a dsh agent tool called md_convert. Failures always carry a stable error code such as E_OCR_EMPTY or E_LEGACY_CONVERT so scripts and downstream tools can triage programmatically. Python and OCR dependencies are installed automatically on first use, and temp files are cleaned up after every run with a signal-hook safety net. This is the plugin to reach for if you are an engineer who needs bulk document-to-Markdown conversion, a DeepSeek Harness user wiring up agent workflows, or a team with a stack of scanned PDFs that need to become LLM-ready text.