fix: convert Tesseract language codes for PaddleOCR in OCRAgent.get_agent()

[Mustafa-Shoukat1]

Fixes #3957 — PaddleOCR receives Tesseract language codes (e.g., eng) instead of its own codes (e.g., en) when using the ocr_only strategy.

Root Cause

All language codes are unconditionally converted to Tesseract format via prepare_languages_for_tesseract() early in the pipeline (pdf.py L344). The hi_res path correctly converts these back to PaddleOCR format in supplement_page_layout_with_ocr(), but the ocr_only path passes them directly through OCRAgent.get_agent() without conversion. The same issue also affects unstructured/metrics/table_structure.py which calls get_agent(language="eng").

Fix

Added language conversion directly in OCRAgent.get_agent() — when PaddleOCR is detected as the configured agent, Tesseract language codes are converted via tesseract_to_paddle_language() before instantiation. This centralizes the fix so all get_agent() callers benefit.

The hi_res path is unaffected since it calls OCRAgent.get_instance() directly (bypassing get_agent()).

Changes

• unstructured/partition/utils/ocr_models/ocr_interface.py: Added tesseract_to_paddle_language import and language conversion in get_agent() when PaddleOCR is configured
• test_unstructured/partition/utils/ocr_models/test_ocr_interface.py: Added 2 test methods (1 direct + 4 parametrized cases) verifying PaddleOCR language conversion through get_agent()
• CHANGELOG.md: Added fix entry under 0.22.18

Testing

• All 15 tests in test_ocr_interface.py pass (including 6 new test cases)
• All 57 tests in test_lang.py pass (unchanged)
• Tested language conversions: eng→en, ara→ar, chi_sim→ch, deu→german

[Mustafa-Shoukat1]

Hi, friendly ping - CI is passing (2/2 checks). This adds PaddleOCR as a lightweight alternative OCR option. Happy to address any feedback or make changes!

[Mustafa-Shoukat1]

Hi, friendly ping - CI is passing (2/2 checks). This adds PaddleOCR as a lightweight alternative OCR option. Happy to address any feedback or make changes!

[Mustafa-Shoukat1]

@badGarnet @cragwolfe - would you have a moment to review this? It fixes the PaddleOCR language code bug reported in #3957 (Tesseract codes like eng are passed to PaddleOCR instead of en). The fix is small (centralizes conversion in get_agent()), CI is passing (2/2), and includes 6 new test cases. Happy to address any feedback!

[Mustafa-Shoukat1]

@badGarnet @cragwolfe - would you have a moment to review this? It fixes the PaddleOCR language code bug reported in #3957 (Tesseract codes like eng are passed to PaddleOCR instead of en). The fix is small (centralizes conversion in get_agent()), CI is passing (2/2), and includes 6 new test cases. Happy to address any feedback!