PRQL actually goes through two intermediary ASTs (PL and RQ) before being compiled into SQL. This is not that well documented yet but have a look at the comment

PRQL actually goes through two intermediary ASTs (PL and RQ) before being compiled into SQL. This is not that well documented yet but have a look at the comments at the top of the lib.rs [0]. Both representations are accessible as json in the Rust and Python APIs. On the roadmap are plans to support DataFrame backends such as Polars, Pandas, Ibis, ... and Substrait.io [1]. The aim of Substrait is similar to what you describe, i.e. to create a universal relational algebra representation that can be executed directly by backend systems without having to go through SQL. [0]: https://github.com/PRQL/prql/blob/main/prql-compiler/src/lib.rs https://github.com/PRQL/prql/blob/main/prql-compiler/src/lib... [1]: https://substrait.io/ https://substrait.io/