Levandor

KB Hungarian translation — prioritized corpus COMPLETE

3 min read

Completion record for the Hungarian translation of the car-repair knowledgebase manuals. The KB manual markdown is now translated to Hungarian for the full prioritized set: 820/820 prioritized pages (956 total .hu.md incl. consolidated). Served read-only under /hu/<page_url> via mkdocs-static-i18n suffix mode, with English fallback for anything untranslated. This finishes the translation half of the i18n design in 2026-07-28-kb-hungarian-i18n-design.

For Agents

  • Delivery: per-page .hu.md siblings next to each English source page, on telep-mainframe. Served at /hu/<page_url> (suffix mode). English is the automatic fallback for any page without a .hu.md.
  • Search is unaffected: the HU mirror is NOT indexed — semantic search stays English-only (bge-large-en). See kb-agent-api.
  • Provenance banner on every translated page: > AI-forditas. A hiteles forras az angol eredeti.
  • State: 820/820 prioritized pages done (956 total .hu.md). A defined skip-list remains (below) and can be added later.

What was translated

820/820 prioritized pages (956 total .hu.md including consolidated articles). Manuals done:

ManualPages
chevy-tracker/geo-tracker-repair434
suzuki-vitara/workshop-1988-1998293
suzuki-vitara/5door-supplement61
suzuki-vitara/supplement-61a4023
suzuki-sidekick/wiring-19969
kick-fix/* how-tosall
consolidated articlesall

Deliberately skipped (low translation value)

Left English on purpose — can be added later if needed:

  • parts-catalogue-89-98
  • body-measurements
  • wiring-diagrams booklet
  • owners-1995
  • kickfix-docs/*

How it was done (the method that survived)

Bulk-translate-then-write loses everything on an API drop — flush per page

The first bulk attempt lost the entire run to a single mid-run API “connection closed” (nothing had been flushed to disk). The fix that made this tractable: each page is written to disk the instant it is translated, then grep-verified. After switching to flush-per-page, three consecutive API “connection closed” drops and a hard “session usage limit” (resets ~3:20am) cost zero completed work — just re-run; a skip-if-.hu.md-exists guard resumes from where it stopped. This crash-safe incremental approach is essential, not optional.

  • Parallelism: session-model FORK subagents each translate a disjoint page range (no overlap).
  • Write mechanism: heredoc / ssh-stdin / scp piping so quotes, $, and backticks in the content survive the shell round-trip to the box.
  • Verification: each page grep-verified after write.
  • Cost: ~6–12k tokens/page — FSM/table-heavy pages at the high end, prose at the low end.

Faithfulness contract

Translate prose + headings. Keep VERBATIM, never reworded or converted:

  • all numbers/units, torque, clearance, resistance, voltage
  • part-numbers, DTC codes, wire-color codes, mileage
  • URLs and every ![](image) reference
  • tables: labels translated, numeric cells verbatim
  • lookup / paint-mark CODES stay in English
  • garbled OCR-noise tables: do not translate

Every page carries the banner: > AI-forditas. A hiteles forras az angol eredeti.

Two open judgment calls to revisit

  • Page 198 — bearing paint-mark colors were translated (Green → Zöld). Revert if they are codes rather than descriptive colors.
  • Page 202 — one garbled OCR TOC cell was dropped.

Status

  • 820/820 prioritized pages translated; 956 total .hu.md including consolidated.
  • Served read-only at /hu/<page_url> (suffix mode), English fallback for the rest.
  • Not indexed for search (search stays English — bge-large-en only).
  • Skip-list (parts catalogue, body measurements, wiring booklet, owners, kickfix-docs) can be added later.
  • Two minor judgment calls flagged for review (pages 198, 202).

Graph View

Backlinks