Sounds like the goal here is to augment search with a sort of query language. One that's more structured than a mere list of keywords?

Ribbing aside, the "rethinking" here sounds a lot like an AI-specific spin on the "bring the code to the data, instead of data to the code" paradigm. I imagine that the same multitenancy caveats of allowing arbitrary user-defined queries against a relational database (execution limits, access controls, etc.) will apply here as well.