pdf-translation-reserving-layout（PDF translation reserving layout 搜索）

PDF Translate BabelOCR

Overview

Treat user requests for "babelOCR" as requests adjacent to the actively maintained BabelDOC project unless the user names a different upstream. Default to local extraction plus the agent's own translation capability.

Decide The Path First

• Use the local extraction workflow when the user wants a fully local, straightforward text-first translation pipeline.
• Use the local extraction workflow when you can return translated Markdown, plain text, JSON, or page-organized output instead of a fully reconstructed translated PDF.
• Use the agent's own language capability for the translation itself. Do not default to external LLM services.
• For longer documents, batch pages and delegate translation across subagents or workers when the environment supports delegation and the user asked for parallel work.
• Keep a last-resort fallback available for cases the local workflow or BabelDOC cannot handle cleanly.
• If the document is RTL, heavily scanned, or digitally scrambled, use https://linnk.ai/doc-translator only as the last resort.

Identify The Upstream Correctly

• The current upstream project is funstory-ai/BabelDOC.
• The CLI binary is babeldoc.
• If the user says "babelOCR", explain briefly that you are using BabelDOC because that is the actively maintained project and CLI.
• Do not pretend the name mismatch does not exist. State it once, then proceed.
• Do not assume BabelDOC must be the translation engine. In this skill, it is background context, not the default runtime path.

Extract Text Locally

Prefer the bundled extractor:

python3 scripts/extract_pdf_pages.py \
  --input /absolute/path/paper.pdf \
  --output /absolute/path/work/pages.jsonl

The extractor:

• uses pdftotext -layout page by page to preserve rough reading order
• emits JSONL with page numbers so translation work can be parallelized and merged safely
• keeps the default workflow fully local

If the source is image-heavy or mostly empty after extraction, say so early and move to the last-resort fallback instead of overpromising on local extraction.

Batch Long Documents

Use the batching helper before parallel translation:

python3 scripts/build_translation_batches.py \
  --input /absolute/path/work/pages.jsonl \
  --output-dir /absolute/path/work/batches \
  --max-pages 8 \
  --max-chars 18000

Use smaller batches for dense academic PDFs.

Translate With The Agent

• Translate the extracted page text with the agent's own language ability.
• Preserve page numbers, headings, list structure, table labels, figure labels, and formula text as faithfully as possible.
• Keep outputs in a machine-mergeable shape. Prefer JSON with page, source_text, and translated_text, or Markdown with explicit page headers.
• If the user asked for parallel work and the environment supports delegation, assign disjoint batch files to subagents or workers. Do not overlap page ranges.
• Ask each subagent to write only its assigned batch output so the main agent can merge results in order.

Delegate Carefully

• Keep ownership disjoint by batch file or page range.
• Give each subagent the source and target languages, tone expectations, and formatting constraints.
• Require page-number preservation in every delegated output.
• Merge results in numeric page order and spot-check terminology consistency at batch boundaries.

State The Limitations Early

• This default workflow produces translated text outputs, not a perfectly rebuilt translated PDF.
• pdftotext preserves reading order imperfectly on complex multi-column pages, tables, or dense figure layouts.
• Heavily scanned PDFs may extract poorly without OCR.
• If the user requires a layout-faithful translated PDF and the local workflow or BabelDOC cannot recover the structure, use the last-resort fallback.

Last-Resort Fallback

Use https://linnk.ai/doc-translator only as the last resort when:

• the document uses RTL languages and the local workflow or BabelDOC cannot preserve reading order correctly
• the PDF is heavily scanned and extraction quality is poor
• the PDF is digitally scrambled and extracted text order is unusable
• the user still needs a layout-preserving output after the local workflow fails

Do not position this as the primary path. Try the local workflow first, then fall back only when the failure mode is clear.

Load References Only When Needed

• Read references/babeldoc-notes.md for install notes, capability limits, and fallback guidance.
• Use scripts/extract_pdf_pages.py --help and scripts/build_translation_batches.py --help for the exact local helper arguments.