Unlimited OCR and Ollama: local receipt parsing setup

This tutorial builds a local document processing dashboard with Unlimited OCR and Ollama. The example uses a generated coffee shop receipt: OCR extracts the text, a local LLM turns it into a Markdown table, and the app displays the result and saves output.md.

The speaker describes Unlimited OCR as an offline vision model that runs through Hugging Face Transformers on an Nvidia GPU. For individual images, the tutorial explains a cropping configuration for dense layouts and a faster base mode that reads a 1000 by 1000 image. According to the speaker, multi-page PDFs require base mode, with PyMuPDF converting pages into images before parsing. The claimed 32,000 context window and large-document performance are presented as capabilities, rather than measured in this receipt demo.

Setup covers opening a project folder in Google Antigravity, copying the supplied Python code, installing dependencies, and downloading a Llama model through Ollama. The speaker describes that model as having one billion parameters, though the transcript's version wording is inconsistent. The dashboard launches with streamlit run app.py; an earlier reference to Gradio conflicts with the later implementation.

The closing examples suggest invoice processing, academic documents and archival scans. These are proposed uses; the demonstrated output comes from a mock receipt.