OCR, translate and structure research PDFs

Works on macOS, Linux and Windows. Allow 8 GB RAM for ordinary papers and 16 GB RAM for large image-heavy PDFs. A local 8B translation model benefits from 8 GB or more unified memory or VRAM.

Component	Role	Price	Link
OCRmyPDF 17.8.1	deskew, OCR and searchable PDF/A or PDF output	Free / 免费MPL-2.0 software; requires Tesseract and system PDF dependencieschecked 2026-08-02	https://github.com/ocrmypdf/OCRmyPDF/releases/tag/v17.8.1
Docling 2.117.0	layout-aware conversion to Markdown and structured JSON	Free / 免费MIT software; local model downloads and compute are separatechecked 2026-08-02	https://github.com/docling-project/docling/releases/tag/v2.117.0
Ollama 0.32.5 with qwen3:8b	local structure-preserving translation	Free / 免费local runtime is free; verify the selected model license for your usechecked 2026-08-02	https://github.com/ollama/ollama/releases/tag/v0.32.5

Create searchable PDF and structured exports

uv venv --python 3.12 .venv
. .venv/bin/activate
uv pip install 'ocrmypdf==17.8.1' 'docling==2.117.0'
mkdir -p work/docling
ocrmypdf --deskew --clean --rotate-pages --output-type pdf input.pdf work/searchable.pdf
docling work/searchable.pdf --to md --to json --output work/docling
ollama pull qwen3:8b
jq -n --rawfile source work/docling/searchable.md '{model:"qwen3:8b",stream:false,options:{temperature:0},prompt:("Translate the following research document into Simplified Chinese. Preserve every Markdown heading, list, table, equation, citation key, URL, figure label and code block. Do not summarize. If a term is ambiguous, keep the original in parentheses. Output Markdown only.\n\n" + $source)}' | curl --fail-with-body -sS 'http://localhost:11434/api/generate' -H 'Content-Type: application/json' -d @- | jq -r '.response' > work/translated.zh.md

Docling batch manifest

input_dir: ./incoming_pdf
output_dir: ./structured_output
outputs:
  - searchable_pdf
  - markdown
  - json
ocr:
  languages: [eng, chi_sim]
translation:
  target_language: zh-CN
quality_gate:
  preserve_headings: true
  preserve_tables: true
  preserve_equations: true
  retain_source_artifacts: true

Known pitfalls

• OCRmyPDF 17.8.1 may emit a regular PDF instead of PDF/A when embedded CJK fonts cannot be safely converted.
• Use --output-type pdf when preservation matters more than archival PDF/A conformance.
• Local LLM translation can alter numbers or citations. Diff tables, equations and reference identifiers before use.
• Docling model downloads make the first run slower and require network access unless pre-cached.

Any machine with Docker and 8 GB RAM. OCR and layout stay local; only extracted Markdown is sent to DeepL if the API step is enabled.

Component	Role	Price	Link
OCRmyPDF Docker image 17.8.1	containerized searchable-PDF generation	Free / 免费local container runtime and storage costs onlychecked 2026-08-02	https://ocrmypdf.readthedocs.io/en/latest/docker.html
Docling 2.117.0	local layout and table extraction	Free / 免费open-source local processingchecked 2026-08-02	https://docling-project.github.io/docling/
DeepL API Developer or Growth v2 text translation API	higher-quality document text translation	Usage-based / 按用量计费unverified exact Growth subscription and overage rate; Developer has a stated 1,000,000-character total allowancechecked 2026-08-02	https://support.deepl.com/hc/en-us/articles/360021200939-DeepL-API-plans

Translate extracted Markdown through DeepL without losing the source

export DEEPL_API_KEY='paste_key_here'
export DEEPL_API_BASE='https://api.deepl.com'
cp work/docling/searchable.md work/source.en.md
test $(wc -c < work/source.en.md) -le 120000 || { echo 'Source exceeds the 128 KiB DeepL request limit; split it by Markdown sections first.' >&2; exit 1; }
jq -n --rawfile text work/source.en.md --arg target_lang ZH '{text:[$text],target_lang:$target_lang}' | curl --fail-with-body -sS "$DEEPL_API_BASE/v2/translate" -H "Authorization: DeepL-Auth-Key $DEEPL_API_KEY" -H 'Content-Type: application/json' -d @- > work/deepl-response.json
jq -e '.translations[0].text | length > 0' work/deepl-response.json >/dev/null
jq -r '.translations[0].text' work/deepl-response.json > work/translated.zh.md
test -s work/translated.zh.md && sha256sum work/searchable.pdf work/source.en.md work/translated.zh.md > work/SHA256SUMS

Known pitfalls

• Legacy API Free and Pro plans can no longer be purchased; new users should verify Developer or Growth endpoint details in their account.
• Sending raw Markdown as text does not guarantee perfect table preservation; validate complex tables manually.
• The exact DeepL Growth charge is intentionally marked unverified.
• Do not send confidential manuscripts to an external API without approval.

No local accelerator required. Requires a Google Cloud project, billing, Cloud Storage and Document AI processor permissions.

Component	Role	Price	Link
Google Cloud Enterprise Document OCR Enterprise Document OCR processor v2	managed OCR with text and layout detection	Usage-based / 按用量计费first 1,000 pages per account-month free, then $1.50 per 1,000 pages up to 5 million pageschecked 2026-08-02	https://cloud.google.com/products/document-ai/pricing
Google Cloud Layout Parser current managed processor	structure extraction and initial chunking	Usage-based / 按用量计费$10 per 1,000 pageschecked 2026-08-02	https://cloud.google.com/products/document-ai/pricing
DeepL API Growth v2 text translation API	translated structured text	Usage-based / 按用量计费unverified fixed subscription, included characters and overage rate; obtain the live quote in the DeepL accountchecked 2026-08-02	https://support.deepl.com/hc/en-us/articles/360021200939-DeepL-API-plans

Cloud pipeline contract

source:
  uri: gs://REPLACE_BUCKET/incoming/input.pdf
ocr:
  processor: enterprise_document_ocr_v2
  native_pdf_parsing: true
  enable_image_quality_scores: true
layout:
  processor: layout_parser
translation:
  provider: deepl_v2
  target_language: ZH
outputs:
  raw_document_ai_json: gs://REPLACE_BUCKET/artifacts/document-ai.json
  source_markdown: gs://REPLACE_BUCKET/artifacts/source.md
  translated_markdown: gs://REPLACE_BUCKET/artifacts/translated.zh.md
validation:
  require_page_count_match: true
  require_nonempty_tables: false
  retain_original: true

Human review checklist for translated structure

Compare the source PDF, OCR JSON and translated Markdown page by page. Report only verifiable defects. Check: page count, heading hierarchy, table row and column counts, equation symbols, units, decimal separators, in-text citations, reference identifiers, figure labels and untranslated fragments. For each defect return source page, artifact path, observed value, expected value and severity. Never invent missing text.

Known pitfalls

• Document AI processors must be created in a supported region and IAM failures can look like empty processing results.
• OCR, layout parsing, translation and storage are billed separately.
• Layout Parser pricing does not include translation.
• Keep raw processor JSON so structure extraction can be rerun without another OCR pass when possible.