[hedora]

I’d expect something like claude 3.5 or chatgpt 4 to do OK at this. (Maybe ocr the receipt, or just send a copy to a multimodal model).

You might be able to use one of the open weight models instead. (Maybe one of the apache 2.0 qwen’s?) Scan a batch, then hand check the results in fifo order. That way you can probably get away with a local gpu.

The hard part will be getting structured output that gnucash likes. I’d try a simple json schema (stick the schema and some example input/output in the context), and then write some code to convert it to a format gnu cash can import.