[smallnix]

Querying an LLM for 'facts' is dangerous. Using some IR technique and incorporating LLMs to gauge relevancy and semantic alignment is a viable approach.