[Prompt Engineering]: Indirectly log the prompt contents
We would like to have a better idea of the actual prompt being sent for inference, but without logging the source code contained in it due to privacy concerns. An idea is to parse the prompt with tree-sitter on model-gateway's side and log the count of symbols. This might give us a better idea of how some prompt transformations affect the contents of the final prompt.
Example |
---|
![]() |