I think to most people it’s pretty obvious you are trying to make the algorithm fit your bias/preconceived ideas.
ChatGPT knows a good answer to the question even without embeddings. But this particular tool can't replicate it, and says e.g. "In summary, these verses highlight sexual ethics and various sins in general, but do not uniformly condemn or condone homosexuality specifically." (which is not wrong, it's just the wrong verses found). (It gives a different summary on different tries)
This is a common problem with embedding search. Obviously, the other traditional techniques would be even worse. But I'd love the systems to be better, and I propose a potential solution, and ask for other ones. I will not be content with your "put up with AI idiosyncrasies and weaknesses, as if they were the real actual conceptual limitation of knowledge" approach. AI has potential to create great UI, but your attitude won't help with that
ChatGPT knows what people in its training corpus commonly say is a good answer.
When you force it to look strictly at the text isolating the influence of popular interpretation, it comes up with a different answer that you like less.
Now, one explanation could be that the embeddings it uses to find relevant text are bad. But there’s another explanation, too...
A software tester walks into a bar.
Runs into a bar.
Crawls into a bar.
Dances into a bar.
Flies into a bar.
Jumps into a bar.
And orders:
a beer.
2 beers.
0 beers.
99999999 beers.
a lizard in a beer glass.
-1 beer.
"qwertyuiop" beers.
Testing complete.
A real customer walks into the bar and asks where the bathroom is.
The bar goes up in flames.
When people agree with the results for the majority of the corpus but cherry-pick “inaccuracies” to justify their bigotry, that’s what I have a problem with.
It sounds like you want to RLHF a model to hate LGBTQ people…
This is not progress.
Its a search of the text by the semantics of the text, not a search of the text by how doctrine has been rationalized (which could be done by an LLM, but wouldn't use a vector DB of the text, rather, it would need sonething like a vector DB of documentation of relevant written justifications of doctrine annotated with scriptural references, and then a simple book-chapter-verse DB of the scriptural text.) These are decidedly different problems, and complaining that something that purports to solve the first doesn’t solve the second is... odd.