It's worth noting that for Amazon in particular there is a special downloader in finance_dl (and corresponding importer in this repo) which would automatically match the credit card transaction and impute Amazon.com as a payee.
The general case of merchants without another source is tricky. Predicting payee is harder than predicting account because 1) there are more of them, 2) if you buy from new places you'll often find yourself out of sample, with new payees that had not appeared before and thus have no training data, 3) it's harder to generalize to a new payee.
An example of this last point is the following:
If you have a training example with narration "Starbucks Coffee XA5757" and account Expenses:Coffee one can see how given a never before seen narration "Local Coffee Shop A97897ADD" the model could generalize and correctly predict Expenses:Coffee. But inferring that the payee is "Local Coffee Shop" on a never before seen example is... either a generative task, or an entity recognition task.

Another approach is to write a standalone script that reads the journal and applies static rules that you would maintain manually. I would advise you try this first because it's not hard to do this from scratch, and may give you insights into solving the problem with ML.

That makes sense. Initially I did try using the beancount example ofx and generic importer source to try the https://github.com/beancount/smart_importer predictor but that didn’t seem to work and also broke all the matching from beancount_import.

I will think a bit on where payee handling might fit. Seems like I could wrap beancount_import.source.ofx and adjust things in a separate SourceResult 🤔