similar data together, extrapolating to create BS, depending on how global that data actually is.

The training data is compressed into weighted parameters, not a petabytes large database, rather like an autocomplete function.

One should incant first that the LLM:
* Answers exclusively based on the search results
* Does not connect the dots, infer or fill in gaps in sources.
* If search results don't explicitly state a fact, return that the search results do not contain the answer.