kordoc ships measured styles for 52 central ministries only. Local governments and public institutions each have their own house style, and there are
hundreds of them, so kordoc cannot contain them all. An agency pack (.kpack) solves this the other way round: you give kordoc a few of your own
agency documents, it builds one file that holds that style, and with that file kordoc writes new documents in the same style. Nothing agency-specific
is added to kordoc, so no agency logo or internal document ends up in the npm package. The Korean manual (agency-pack.md) is the reference.
kordoc pack build report1.hwpx report2.hwpx report3.hwpx --name Gwangyang -o gwangyang.kpack # same kind of document, 3–10 samples
kordoc pack info gwangyang.kpack --guide # summary + writing guide
kordoc generate new.md --pack gwangyang.kpack -o new.hwpx # manuscript written per the guide
kordoc pack build press1.hwpx press2.hwpx --name Gwangyang --kind press --into gwangyang.kpack # add another document kindAPI: buildAgencyPack(name, samples, into?), loadAgencyPack(buf), packTemplate(pack, kind?) → markdownToHwpx(md, { template: packTemplate(pack) }).
MCP: build_agency_pack, agency_pack_info, and generate_document with pack_path.
A pack is a zip with no code: pack.json (name, kinds, representative, roles, per-sample agreement, self-check, leftover text, image count, sample
SHA-256 fingerprints), kinds/<kind>.hwpx (the representative sample with its original text stripped — styles, frames and logos only) and guide.md.
- Representative — per-role (title box, contact line, summary box, chapter band, □·ㅇ·-·※, body) style votes over font, size, bold, alignment and line spacing pick six candidates; they are re-ranked by cross-reproduction: regenerate the other samples' text with each candidate and compare formats.
- Stripping — a skeleton manuscript that uses each role once is generated with the representative; author/date metadata is cleared, names and phone numbers left in approval-line cells become ○, and department prefixes before job titles are removed.
- Self-check — the stripped template is re-read; if roles or styles differ from the original, the next candidate is used, otherwise a warning is kept.
- Loading treats a pack as untrusted: zip size and entry limits, only
pack.json,guide.md,kinds/*.hwpx, and a format version check.
Same scorer as the generation fidelity bench (bench/gen-repro.mjs). Every number uses a pack built without the document being scored.
| Set | Manuscript | Plain report preset | Agency pack | Upper bound (own template) | Font/size/bold (preset → pack) |
|---|---|---|---|---|---|
| Seoul policy plans, 5 of 16 with manuscripts | hand-transcribed | 68.3% | 74.3% | 80.0% | 53.5% → 75.3% |
| Gwangyang report style 1, 9 docs (from proposal #173) | same | 77.1% | 94.5% | 95.4% | 38.8% → 85.9% |
| Gwangyang website notices, 4 docs (same form) | parsed by kordoc | 58.3% | 71.1% | 74.4% | 24.8% → 55.2% |
How to read these: 100% is out of reach with any single template — authors hand-tune line and paragraph spacing per paragraph, so even a document's own template scores about 80%. The Seoul and Gwangyang-notice sets are human-written and closest to real use. The Gwangyang style-1 ground truth was produced by a generator (uniform formatting) and its generator-only components (callout boxes, rounded boxes, table-title bands) were removed from the manuscripts, so real Gwangyang reports will score lower. Sets are small (4–9 documents).
Stripping is lossless for all 144 Seoul template × manuscript cases (the one template whose skeleton-stripped form drifts falls back to the masked form, which is byte-identical in output). The privacy check
(names, phone numbers and 8-character body fragments of the sample must not appear in the pack) passes for all 16 Seoul and 9 Gwangyang samples.
Reproduce with node bench/pack-repro.mjs [--dir=<folder>] [--privacy].
Kind names are free-form (report by default, press, notice, …); they only select a template. Lines that start with a number and a period
(2026. 3. 1.) are read as Markdown ordered lists — escape the period (2026\. 3. 1.); generation warns about it. Template-mode limits apply (no images, footnotes or inline math; no cover or table of contents; one body section). Documents that use only numeric
hierarchy (1. → 1) → ①) benefit little. Single-table forms should use fill_form/patch_document. HWP 5.x samples must be saved as HWPX first.
Packs carry no original body text, author data or approval-line names, but they do carry logos, headers and footers — check before sharing.