Raw LLM Responses

I got to similar conclusions I find by vectorial similarly (not sure it is still cosine the algo I chose but essentially the same thing) and the LLM then just accept / reject each of them. Actually the LLM don't accept or reject them, he has to qualify each result on several axis.  Like in your case it would be something "relevance determine how relevant it is [critical,relevant, tangential, irrelevant], how applicable it is [applicable, wrong geographic... Also adding axis allowing to class "irrelevant" like if you have old passage not applicable anymore: [Historical, supersede, in vigor] Usually 3-4 positive axis and adding negative axis as "magnets" for specific false positives. By code I then filter on those axis via an empirical scoring mécanism.