ocr_min_words: 10 # Pages with fewer words from text layer → OCR fallback batch_size: 32 # Pages embedded per Ollama call (tune to your GPU VRAM)