Language models are great at turning those statements into Beancount postings and fixing errors, but the local ones not so much yet.
Language models are great at turning those statements into Beancount postings and fixing errors, but the local ones not so much yet.
Why would someone need this? How do you use this?
Check out https://plaintextaccounting.org/
(I don't use beancount but I use GnuCash which is a very similar concept.)
The double-entry part is a fanciness for checking that all money in (paychecks) equals all money out (spending, taxes). You can selectively disable it by “padding” balances.
I have it on for my checking account and it catches duplicate postings for transfers between my Venmo and my checking account. I can also put the initial/final balances from the bank statements as “guards” to verify that it all adds up.
This sort of structure is really helpful for a human or language model. You can break the translation work down into subunits with fast correctness checks.