Query OCR

Use This When

• You want one OCR worker with pluggable backends
• Switch between lightweight OCR (Tesseract, EasyOCR) and VLMs (MiniCPM, LLaVA)

What It Does

• Accepts an Image and runs OCR using the selected backend
• Backends:
  • tesseract (CPU, classic OCR)
  • easyocr (CPU/GPU, detection+recognition)
  • surya (Surya OCR; downloads model weights on first use)
  • minicpm (VLM via Transformers; prompt-driven)
  • llava (VLM; visual QA style)
• Lazily initializes heavy models and reuses them across calls

Key Config

• backend: tesseract | easyocr | rapidocr | paddleocr | trocr | doctr | minicpm | llava | gotocr | surya
• use_gpu: enable GPU acceleration where supported (defaults to true; falls back to CPU if not available)
• language: language code(s). For Tesseract use ISO3 (e.g., eng), for EasyOCR use 2-letter (e.g., en). eng is auto-mapped to en.
• prompt: instructions for VLM backends
• Tesseract tuning: tesseract_oem, tesseract_psm
• Surya: surya_task_name, surya_output_format, surya_disable_math, surya_sort_lines, surya_separator, surya_recognition_batch_size, surya_detector_batch_size, surya_foundation_chunk_size, surya_foundation_max_tokens, surya_foundation_model_quantize, surya_disable_tqdm
• Hugging Face model ids:
  • trocr_model, minicpm_model, llava_model, gotocr_model
• LLaVA: num_tokens
• GOT-OCR: gotocr_mode, gotocr_crop, gotocr_box, gotocr_color
• RapidOCR: rapidocr_text_only
• PaddleOCR: paddle_lang
• PaddleX cache (PaddleOCR): paddlex_models, paddlex_revision
• TrOCR: trocr_use_rapid_det
• docTR: doctr_det_arch, doctr_reco_arch

Notes

• Tesseract requires the Tesseract binary installed in the runtime image.
• EasyOCR uses PyTorch and OpenCV; GPU is optional via easyocr_gpu.
• Surya downloads model weights on first use (requires network access in the runtime environment).
• VLM backends are much heavier; GPU recommended.

Build

• Per repo rule, build with: ppl compile --non-interactive from this component directory (requires elevated permissions).