Last December, I said that I didn't have a use for LLM based generative AI in ticketing.
Well, I found one.
At the Ticketing Professionals Conference 2024, in the wake of his keynote presentation that I didn't actually go and see, I was chatting with Rob Williams, late of AudienceView, about my scepticism and his enthusiasm for our coming AI overlords. And one of the things that he said really stuck: that LLMs "work" by (in some way) translating an English sentence into a set of statistical weights in some internal private language that represents the meaning of the sentence, and then it can do stuff with this, and then translate the answer back into English... but it can translate the answer back into French or German or computer code equally well.
And this made me realise that if I gave an LLM my database schema - but none of the actual data - then it could translate questions into SQL, which I could then run on my database, and get useful, accurate answers. The answers would be SQL - auditable and checkable and repeatable, not just some numbers that it had made up and had no way of showing the working. If the SQL was valid, it would run on the database, if it was invalid, it wouldn't. There would be much less danger of the LLM "hallucinating" an answer and making stuff up and getting it wrong. And there would be no danger of private data being sent to an AI and being used as training data and leaking out in some way.
So that's what I built:
No comments:
Post a Comment