Is retrieval necessary/sufficient to solve “hallucinations”?

Hi everyone,

"Hallucinations" seem to be one of the major blockers to the adoption of LLMs, especially in enterprise settings where seemingly plausible but false information/decisions can be disastrous.

I am wondering whether or not current LLMs, like GPT4 or Llama 2 70b have reached the reasoning capabilities to be able not to hallucinate when fed the proper information.

One can see a LLMs as college / high school students that have some basic reasoning and knowledge but might need to be nudged to answer business / scientific questions.

What is your opinion on the topic? Is there for instance, a database of hallucinations, so that one could test that if the model would have answered properly if the right context was fed in the prompt?