You are correct. For some data sources like parquet it includes some metadata that helps with this, but it's not as robust at databases I dont think. And of course, cvs have no metadata (I guess a header row.)

The actually specification for how to efficiently store tabular data in memory that also permits quick execution of filtering, pivoting, i.e. all the transformations you need...is called apache arrow. It is the backend of polars and is also a non-default backend of pandas. The complexity of the format I'm unfamiliar with.