OCR Language Support#

Docling's OCR subsystem supports multiple engines — RapidOCR, EasyOCR, Tesseract (CLI and tesserocr), macOS Vision, and others — each with its own language configuration model. Language is set on the engine-specific options class passed to PdfPipelineOptions.ocr_options 1.

RapidOCR: Bundled Model Sets#

RapidOCR is the most constrained engine: it bundles exactly three model sets — chinese, english, and latin — rather than supporting arbitrary language codes directly 2.

Language resolution happens in _resolve_rapidocr_language(), which maps user-supplied codes (ISO 639-1, ISO 639-2, and plain English names) to one of these three sets. Key rules:

• en/eng/english → english model set
• de, fr, es, it, pt, nl, and ~20 other European languages → latin model set 3
• zh/zho/chi/chinese → chinese model set
• If no matching group is found, a warning is emitted and chinese is used as the default — which silently drops inter-word spaces in Latin-script text 4
• If multiple requested languages span different model sets (e.g., ["en", "zh"]), only the first resolved set is used, with a warning 5
• The latin set subsumes english: requesting ["en", "fr"] resolves to latin, not a conflict 6

Language codes are normalized before lookup — "en-US" becomes "en", "fr_CH" becomes "fr" 7.

⚠️ Latin model not downloaded by default. The download_models() utility only downloads chinese and english by default. The latin set must be explicitly requested or pre-loaded 8. This gap caused a real regression when the docling-serve rocm72 image omitted the Latin model 9.

When artifacts_path is set, each model file resolves to <artifacts_path>/RapidOcr/<relative-path> 10. When no artifacts_path is set and paths are not pinned, language parameters are passed through RapidOCR's internal LangDet/LangRec typings API via _rapidocr_lang_type_params().

Key file: rapid_ocr_model.py

EasyOCR: Delegated Multi-Language Support#

EasyOCR supports 80+ languages and accepts a lang list of ISO 639-1 codes directly. The default is ["fr", "de", "es", "en"] 11. Language codes are passed verbatim to easyocr.Reader(lang_list=...) 12; Docling does no pre-validation — EasyOCR itself handles language resolution and model downloads.

Models for default languages (english_g2, latin_g2) are pre-downloaded via EasyOcrModel.download_models(). For offline use, set artifacts_path; this disables auto-download and points the reader to <artifacts_path>/EasyOcr/ 13.

Key file: easyocr_model.py

Tesseract (CLI and tesserocr): ISO 639-2 Codes + Strict Validation#

Both Tesseract options use 3-letter ISO 639-2 codes (e.g., eng, fra, deu). The default for both is ["fra", "deu", "spa", "eng"] 1415. Languages must have corresponding Tesseract data files installed in TESSDATA_PREFIX (or the path set via TesseractCliOcrOptions.path / TesseractOcrOptions.path).

Tesseract CLI validates language codes via _sanitize_lang(), which enforces the pattern ^[a-zA-Z0-9_/][a-zA-Z0-9_/+-]*$ before any subprocess call. This was added as a security hardening to prevent command injection 16. Multiple languages are joined with + and passed as a single -l flag. All subprocess calls use shell=False.

tesserocr (Python bindings) concatenates language codes with + and passes them directly to PyTessBaseAPI.

macOS Vision (OcrMacOptions)#

Uses locale codes in language-REGION format (e.g., en-US, fr-FR). Default: ["fr-FR", "de-DE", "es-ES", "en-US"] 17. Delegated entirely to Apple's Vision framework; no Docling-level language validation.

Quick Reference#

Key Source Files#