I found the only way to make this work reliably is to treat the LLM as a fallible component inside a state machine rather than the controller. I've been using LangGraph to enforce structured outputs and run validation checks before any side effects happen. If the output doesn't match the schema or business logic it just retries or halts. It seems like a lot of boilerplate initially but it is necessary if you want to trust the system with actual invoices.
If you have to manually validate everything then what did you save by using an LLM? DIY and know it will work the first time.
I found the only way to make this work reliably is to treat the LLM as a fallible component inside a state machine rather than the controller. I've been using LangGraph to enforce structured outputs and run validation checks before any side effects happen. If the output doesn't match the schema or business logic it just retries or halts. It seems like a lot of boilerplate initially but it is necessary if you want to trust the system with actual invoices.
So when this issues a valid but garbage invoices, then what?
https://en.wikipedia.org/wiki/Bernoulli_trial
I remember studying this in uni lol. How do you use it?