The generate tool is extremely useful to build fast and more reliably some transform blocks that have repetitive logic.
One potential improvement of the platform would be to not only generate from scratch, but feed one or several block themselves as input to the LLM for it to change the code (potentially on a branch to benefit from a “diff” like workflow to detect hallucinations)?
Thanks!
Is the use case here simply to almost copy a set of logic but where each copy has some minor differences that only an llm could fit?
Not copy, but edit: let’s say I have a an if/else statement:
if value <10 then “a”
elif value <100 then “b”
elif value <1000 then “c”
…
In pipeline builder it would take me quite a few clicks if I wanted to update 10, 100, etc. to 20, 200, etc.
This would probably be an easy job for an LLM to take adapt the json serialization from a prompt “Double all the values of the if/else statements in this transform block” and save a bunch of time (especially for users that wouldn’t copy the serialized json, edit in a IDE then paste back the new transform and adjust the links manually to use the new version).