I've been curious to see the process for selecting relevant context from a long conversation. has anyone reverse engineered what that looks like? how is the conversion history pruned, and how is the latest state of a file represented?
We plan to continue investigating how it works (+ optimize the models and prompts using TensorZero).