hwpx2md 0.1.0

Build LLM-friendly Markdown artifact bundles from HWPX and HWP documents.

hwpx2md

Convert Korean .hwpx or .hwp documents into an LLM-readable Markdown artifact bundle.

Use this when a model or script needs the document body, memos/comments, tables, and extracted assets in plain files instead of a binary office document.

Quick start

Run directly from PyPI with uvx:

uvx hwpx2md "input.hwpx" -o "input_bundle" --overwrite
uvx hwpx2md "input.hwp" -o "input_bundle" --overwrite

If -o is omitted, hwpx2md writes the bundle next to the source file as <input_stem>_llm_bundle.

Use quotes around Windows paths, Korean filenames, and paths containing spaces.

What you get

After conversion, read the files in this order:

1. document.md: main body text. Memo, table, and asset references are linked inline.
2. memos.md: all memo/comment text. Each memo includes the document anchor, paragraph text, and nearby context.
3. chunks.jsonl: retrieval-friendly chunks. Each JSON line contains text plus related memo, table, and asset IDs.
4. manifest.json: machine-readable inventory with backend, counts, warnings, memo metadata, table metadata, and asset metadata.
5. tables/: one set of artifacts per table: full Markdown, CSV, compact Markdown, and cell JSON.
6. assets/: copied embedded or preview assets when the backend exposes them.

The CLI prints a JSON summary with the output directory, backend, counts, files to read first, and warning codes:

{
  "output": "C:\\work\\input_bundle",
  "format": "hwpx",
  "backend": "native-hwpx",
  "counts": {
    "sections": 1,
    "paragraphs": 96,
    "tables": 5,
    "memos": 2,
    "assets": 1,
    "asset_refs": 0,
    "chunks": 2,
    "warnings": 1
  },
  "read_first": [
    "C:\\work\\input_bundle\\document.md",
    "C:\\work\\input_bundle\\memos.md",
    "C:\\work\\input_bundle\\chunks.jsonl",
    "C:\\work\\input_bundle\\manifest.json"
  ],
  "warnings": ["nested_tables_folded"]
}

Backend fidelity

.hwpx uses the native XML backend. This is the highest-fidelity path. It preserves memo IDs, authors, timestamps when present, document anchors, table cell spans, and exposed assets.

.hwp uses hwp-hwpx-parser. This path does not require Hancom Office or Windows COM automation. It preserves body text, tables, memo text, and [MEMO:N] anchors, but the backend may not expose author, timestamp, exact memo IDs, image references, or table span metadata. Check manifest.json warnings for the exact limitations seen in a converted file.

For the best memo and table metadata, prefer .hwpx when you can. Use .hwp when the original binary file is all you have and text/memo extraction is more important than perfect layout metadata.

Commands

Convert HWPX:

uvx hwpx2md "proposal.hwpx" -o "proposal_bundle" --overwrite

Convert HWP:

uvx hwpx2md "proposal.hwp" -o "proposal_bundle" --overwrite

Show the self-contained CLI guide:

uvx hwpx2md --help

Use the package from a local checkout:

cd hwpx2md
uv run hwpx2md "..\email\sample.hwpx" -o "..\.backup\sample_bundle" --overwrite

Python API

from pathlib import Path
from hwpx2md import bundle_document

manifest = bundle_document(
    Path("input.hwpx"),
    output_dir=Path("input_bundle"),
    overwrite=True,
)

print(manifest["counts"])

Use bundle_hwpx(...) if you want to accept only .hwpx, or bundle_hwp(...) if you want to accept only .hwp.