Nutlope/llama-ocr ? reverse-engineered prompt

Build me a small TypeScript library that takes an image and turns it into clean markdown using Llama 3.2 Vision through Together AI.

I want to be able to call something like ocr({ filePath, apiKey }) for a local image, and also pass a remote image URL. It should return markdown text that keeps the document structure as well as it can, like headings, paragraphs, lists, tables, and simple formatting. Please make the API easy to use from Node, with sensible defaults, but let me choose the model too, including a free option and the faster paid ones if needed.

Keep it simple and production ready, with good error handling and clear docs in the README. If there are tests, cover the main happy path and a couple of failure cases. If you need to check current Together AI docs or example usage, look them up online.