3 ms·

I’d use the LLM as an intake layer, not the database: raw sources stay untouched, extracted claims become structured rows with citations and open questions.