I felt they would just "cast the net wide" with a quick search-collect at scale, then load in the LLMs own training on to the top, and reports I generated were giving me hallucinated content.

I wanted something more - collect-evaluate-decide loop to iterate through discoveries and actively seek out diverse sources.