When generating pipelines to handle polymorphic data, the BI-Connector creates a lot of fields that are not necessarily needed for the final output, as a simplified example:

{$addFields: {x: ..., y:..., z:...}}
{$project{OUT: "$x"}}

The code for computing y and z, which can be quite expensive, is totally unneeded. This can generalize to any number of stages. Essentially, any fields removed by $projects can be removed from computations proceeding said $project, transitively, e.g.:

{$addFields: {x: ..., y:..., z:...}}
{$addFields: {a: {$add: ["$x", "$y"]}, b: {$add: ["$y", "$z"]}}}
{$project{OUT: "$b"}}

Here, we can remove the computation for a, and then, transitively, x. This has the possibility to drastically improve many types of queries, but will require a field level dependency tracker. This will see benefit for any generated code, not just from the BI-Connector, so it makes more sense to be done within the server.