curious if you tried local LLM models for OCR, like a Gemma4, or your volume is too much for that